Challenges in evaluating AI systems / Anthropic

Get Joe Sispak’s thoughts
The “Moody’s” of AI: new tech providing an unbiased evaluation of new generative AI models
Patronus AI Creates an LLM Evaluation Tool for Regulated Industries

Talk to https://www.linkedin.com/in/tbehailu/ about Patronus and Arize before having an in-person conversation with Patronus.
Make sure you get the intro from Gelila first and explain your thought process
Dive deep into everything and ask the real nitty gritty questions

• Automated evaluation platform for LLMs
• Evaluation run to measure model performance
• Adversarial testing sets made to break models
• Generate new adversarial test sets
• Benchmarking performance to other models

Checking if they break in the real world scenario towards regulated industries where there is little tolerance for errors.
“a trusted third party when it comes to evaluating models.”

Rigorous Evaluation

Score + Test + Benchmark

What are my thoughts on the issue being solved:

LLMs are nondeterministic — they’re not guaranteed to produce the same output every time for the same input.
That means that companies will need to do more rigorous testing to make sure they’re operating correctly, not going off-topic, and providing reliable results.
Generate adversarial test cases, monitor hallucinations, and detect PII and other unexpected and unsafe behavior. Customers use Patronus AI to detect LLM mistakes at scale and deploy AI products safely and confidently.
LLMs in current form won’t generate any revenue. Enterprises need much more tech & domain training to align to business.
Good thing from our enterprise clients that they are willing to pay for solution as long as it aligns to business goals
Sideways – test normal modes
High-priority harm areas: self-harm, physical harm, illegal items, fraud, and child abuse.

Questions

• What are traditional methods of testing
  • Academic benchmarks
  • Human evaluations
• Why would someone use us instead of public datasets?

There’s obviously a clear problem that needs fixing but I worry that it’s a hyper-competitive space. I’d love to hear about how they’re trying to make themselves distinct by directly working with observable surfaces (Mongo/Hugging Face…).
I was looking at Arize and others that seemed slightly more advanced but playing directly on the same field.

From what I can tell, they’re relying on the fact that LLMs are nondeterministic — they’re not guaranteed to produce the same output every time for the same input. That means that companies will need to do more rigorous testing to make sure they’re operating correctly, not going off-topic, and providing reliable results. Basically Score + Test + Benchmark… Good thing is they’re targeting enterprise clients that are willing to pay for a solution as long as it aligns to business goals. Eventually become “a trusted third party when it comes to evaluating models.” There’s obviously a clear problem that needs fixing but I worry that it’s a hyper-competitive space. I’d love to hear about how they’re trying to make themselves distinct by directly working with observable surfaces (Mongo/Hugging Face…). I was looking at Arize and others that seemed slightly more advanced but playing directly on the same field. What Patronus seems to be building is the key features in a stack of a bunch of different processes that are required from a company starting to prep data to getting the final results. Obv it’s too early, and I’m in no authority, but I’d assume most of this would be done in one platform that can plug and play different things and the platforms that facilitate that (similar to how you can plug and play with the base models themselves) will eventually build out an evaluation tool. And if there’s a better evaluator, it becomes a plug that can be added more than its own platform. Whether they even plan on moving to the entire space of monitor, finetune, observe, test, evaluate or just master the evaluate section and eventually be an option for the evaluation as they’re the best… (I’d also assume the best plug-and-play evaluators would become super niche industry specific) Nonetheless I would love to hear from him and am interested in talking about the position you mentioned and getting to learn more about how they’re trying to position themselves.

Evals & Leaderboards - what’s ultimately the goal?

What is the goal for evaluation and leaderboards?

Starting from goals for the outcomes we’re seeing to have a nuanced overview.
There are a number of ways to evaluate a model (coding, reasoning, etc.) but the line of sight from something like MMLU or Hellaswag to what the downstream consumer (i.e., the developer) wants in terms of application performance is unclear and certainly non-linear.
Having evals that accurately represent success in the application environment gives model developers something to shoot for—and god knows, when we get a good benchmark to optimize toward, we as a community go nuts!!

Moreover, while it’s still very early, there are signs that we can more confidently start from desired capabilities and adjust the data mixture at various stages of the model development process to align our models more toward how we want them to behave in the eventual application. Whether we need to go all the way back to pre-training is still a good question but we certainly know we can do a lot at the CPT and SFT stages to get models to behave in our desired ways (e.g., speak languages other than English).

Back to goals - what are the goals of evals and leaderboards?

The answer is clearly dependent on who you ask. It can be a combination of wanting to find the best model across performance, type (i.e., base, chat, multimodal, different languages, etc.), latency, and cost—or certainly a local optima across all of them depending on who the developer is and what his or her application goal is, level of sophistication, desired level of customization, etc.

As a model developer (or at least one that likes to play with the models out there) I tend to use Helm or OpenLLMLeaderboard, sort on the provided metrics, see which models rise to the top and which can fit onto a Colab instance so I can just quickly load something up and hack a prototype. My job to be done (JTBD) or goal is pretty straightforward. But for many others, not so much.

Let’s walk through some of the leaderboards, some old some new, and what a goal could be for each. Note this is not exhaustive and is a curated set of leaderboards from the articles I read with my personal perspective on each.

(1) Open LLM Leaderboard or HELM

Maintainer: Hugging Face & Stanford respectively

Summary:
A sortable set of models, open and closed, evaluated on mostly academic benchmarks. Covers core scenarios such as Q&A, MMLU (Massive Multitask Language Understanding), MATH, GSM8K (Grade School Math), LegalBench, MedQA, WMT 2014 along with other benchmarks. These are largely open-source evals that have been developed over the years by various researchers.

Developer goal:
General AI developers looking for models to build on will come here first, understand per a given size which models perform the best, and then leverage the result for their research, an application, fine-tune for a custom application. One thing that’s clear, however, is that there is a stage of disillusionment that the community is in when it comes to actual performance of models and how each rank on academic evals. This has driven the development of other leaderboards such as Chatbot Arena (see later).

(2) Hallucinations Leaderboard

Maintainer: Hugging Face

Summary:
It evaluates the propensity for hallucination in LLMs across a diverse array of tasks, including Closed-book Open-domain QA, Summarization, Reading Comprehension, Instruction Following, Fact-Checking, Hallucination Detection, and Self-Consistency. The evaluation encompasses a wide range of datasets offering a comprehensive assessment of each model’s performance in generating accurate and contextually relevant content.

Developer goal:
Have a starting point to understand which models hallucinate the least. There is some overlap with other available leaderboards here and, given without RAG or search integrated, these models will always hallucinate; this leaderboard is less interesting for developers.

(3) Chatbot Arena / LMSys

Maintainer: Together AI, UC-Berkeley, Stanford, and several others.

Summary:
A benchmark platform for large language models (LLMs) that features anonymous, randomized battles in a crowdsourced manner. Some, like Karpathy, find it to be the best test of a model’s performance as the arena is much more dynamic and compares models head to head.

Developer goal:
Understand how models will perform in a real environment that challenges in ways that standard datasets/prompts can’t. Like #1 above, this may influence the developer’s starting point for further work.

(4) The MTEB leaderboard

Maintainer: Hugging Face & Cohere

Summary:
MTEB or “Massive Text Embedding Benchmark” contains 8 embedding tasks covering a total of 58 datasets and 112 languages. Through the benchmarking of 33 models on MTEB, establish the most comprehensive benchmark of text embeddings to date.

Developer goal:
Understand which models could perform best in applications where retrieval augmented generation (RAG) is used.

(5) Artificial Analysis

Maintainer: Independent / Startup

Summary:
Artificial Analysis provides benchmarks & information to support developers, customers, researchers, and other users of AI models to make informed decisions in choosing which AI model to use for a given task, and which hosting provider to use to access the model.

Developer goal:
Understand cost, quality, and speed trade-offs across models and model providers (e.g., OpenAI, Microsoft Azure, Together.ai, Mistral, Google, Anthropic, Amazon Bedrock, Perplexity, Fireworks, Lepton, and Deepinfra).

(6) Martian’s Provider Leaderboard

Maintainer: Martian (startup)

Summary:
Martian’s provider leaderboard collects metrics daily and tracks them over time to evaluate the performance of LLM inference providers on common LLMs.

Developer goal:
Understand quickly cost vs. rate limits vs. throughput for a variety of LLM providers and optimize for his or her application.

(7) Enterprise Scenarios Leaderboard

Maintainer: Patronus, Hugging Face

Summary:
Enterprise Scenarios leaderboard evaluates the performance of language models on real-world enterprise use cases (at present 6 benchmarks - Finance, Legal, etc.).

• FinanceBench (accuracy not retrieval)
• Legal Confidentiality (reason over legal cases)
• Writing Prompts (engagingness)
• Customer Support Dialogue (relevance)
• Toxic Prompts (generation)
• Enterprise PII (business-sensitive info generation)

Developer goal:
This leaderboard is still very much nascent as you can see from the lack of public models competing. Overall, though, this set of evals would provide developers with an idea of which base models to use as a starting point for a given task. Again, very, very early.

(8) ToolBench

Maintainer: SambaNova Systems

Summary:
A tool manipulation benchmark consisting of diverse software tools for real-world tasks.

Developer goal:
Illustrate which available models perform best for action generation as described in natural language. Note it seems like this project is also a great resource for broader developers for how to implement a particular tool usage (i.e., plug-in) for their model.

Key takeaways:

We are in that stage of the market where it’s an absolute gold rush—if an entity controls the underlying benchmark, this can influence the direction of the market. This is why startups that are dependent on performance/LLM efficiency are pushing on their own benchmarks. If these benchmarks take hold, it will force competitors to follow.

We are absolutely in the early innings and nothing has been decided. If you can dream of a use case or requirement, you could probably define an eval and leaderboard.

Right now Meta is a passive player and, beyond the CybersecEval we released late last year, just leverages the currently available evals and leaderboards.

Lastly, as Alan Kay the PARC researcher said, “The best way to predict the future is to invent it…” - we are now on the path to AGI and it’s time for us to define the future. Buckle up!!

Cheers!

How do you evaluate models (coding, reasoning, etc.)

MMLU or Hellaswag to downstream consumer (i.e., the developer) wants in terms of application performance.
You can just look at the provided metrics, see which models rank higher, and use the one that goes into your code easily to prototype something. E.g.:

• Holistic Evaluation of Language Models (HELM)
• Open LLM Leaderboard - a Hugging Face Space by HuggingFaceH4

What’s the true need of evaluations?

Evals are the new PRD?

1. Find the best model across performance types (chat, multimodal, switching languages, etc.)
2. Finding the optimal range for latency and cost
3. Comparing levels of sophistication or customization levels

Having evals that accurately represent success in the application environment gives model developers something to shoot for. When we get a good benchmark to optimize toward, without a really strong POV on where we want to be going, it is impossible to get there.

Outside of the common ones like cost, safety, quality [case evals], and latency.

An important one of those challenges relates to how we can reasonably evaluate the capabilities of these apparently highly-skilled models. How do we quantify what they can and cannot do? How can we measure how biased they are, or what is the risk they will express toxic or harmful language? And what are useful automatic signals that can be used to improve them even further?

We can work backwards from desired capabilities and adjust the data mixture at various stages of the model development process to align our models towards how we want them to behave in the eventual application.

Do we even need to go all the way back to pre-training or not is a different question, but we know that there’s a lot we can do even at the CPT and SFT stages to make our models behave in our desired ways (change languages, etc.)

IMO, eval is the largest challenge we face. If we can’t show that the models are working/achieving our goals for them, what are we even doing? The crux of the issue is that it’s a challenging thing to scope out what a “general purpose” system ought to be able to do.

Learnings from Current Instances (which are listed in detail below)

• If an entity controls the underlying benchmark, this can influence the direction of the market. Every startup whose product is dependent on performance/LLM efficiency is pushing on its own benchmarks. Once a benchmark becomes the norm, competitors are forced to follow. Thus it’s a gold rush to control the underlying benchmark.
• The best way to predict the future is to invent it…

It’s very early and nothing has been decided. So if you can see a desired capability or requirement, you can define an evaluation and leaderboard.

Every LLM org is going to have to define their version of AGI and build the evals towards that. E.g., OAI is biased towards a single agent, Google is benching on “search” for general intelligence.

Evaluating a model taken to an extreme limit is as complex as coming up with an intelligent model. Back in the days (lol) we got away with easier evals because the dimensionality and decision of taste were very tractable and were often deterministic.

With the incredible advances in foundation models to conduct a wide range of tasks, it becomes more and more challenging to evaluate them and understand their limitations: if a model could potentially execute infinitely many tasks, it is impossible to do a proper evaluation of all the tasks it could do.

Evaluations are subject to human preferences which vary in small populations. So the more models we have, differentiation is preference on human evaluation preferences.

Maybe the head-to-head Chatbot Arena might work to mitigate that, but it can be bottlenecked by the number of people that vote.

Plus head-to-head battles won’t pinpoint to the source of difference in scores.

Who are the stakeholders in AI Evaluations?

• Productionisers: People making models for a purpose that want to make sure it performs robustly, safely, and responsibly.
  • Hard to do automatically as it’s a higher skill level focused. Human evaluations are used and are expensive.
  • For productionisers, it is less relevant to score well on benchmarks that test specific skills in isolation, or in a context different from the one the model will be deployed in.
• Researchers: The opposite of what I said for productionisers.

Dimensions in Choosing Evals

• Generation vs Ranking
  • Multiple choice (ranking or classification tasks) or direct evaluation of generated text (free-form).
  • Ranking can be multiple choice, yes/no, etc.
  • Can be automatic and cheap and useful in model development to run continuously during iterations.
  • Disadvantage is resemblance of scenario being modeled, but not the exact scenario; plus data contamination of the evaluations being part of the training data.
  • Direct open-ended evaluations are harder to do automatically… especially for non-deterministic answers. Human evaluation is the gold standard and suffers from poor replicability and disagreement around criteria and measurement and quality judgment from workers.

Prompt Handling

• The variation of the prompt used can lead to the models giving out varied outcomes from the slightest arbitrary choices. E.g., choice of wording, number of prompts, and the way prompts are formulated and formatted.
• Maybe looking through the best-case scenarios before computing evals to choose prompts that models respond well towards.
• Prompt engineering for evals depends on the stakeholders and situations. For builders, re-engineering the prompt to choose the best performer can lead to seeing the upper-bound of the performance task and potentially re-routing similar prompts to the best performing one. However, for performance researchers aiming to achieve AGI, prompt engineering scores are less informative as variations show how far off we are from corrective intelligence.
• Also, if a model responds unexpectedly to prompt-format changes in a relatively simple ranking test, what does that imply for the reliability of scores on tests for bias or toxicity?

Handling Outputs

• Especially to ranking ones. It’s going to become increasingly easy to assign higher eval scores to models that are eval-instruction-tuned. So a simple yes/no might not be the move… Maybe looking at the assigned probabilities to each option (either in its decision or by passing the prompt multiple times and seeing the probability of choice for each option) could be useful.

Data Contamination

• Relationship between training and evaluation data.
• If you’re an AGI researcher, this is a no-go. Testing on stuff the model might have already seen counteracts the test and is not indicative of hyper-generalization and pattern matching.
• There was a reported score of accuracy of GPT-4 on HellaSwag evals that found 36% of reasoning questions were wrong, mislabeled, or incomprehensible examples. HellaSwag or HellaBad? 36% of this popular LLM benchmark contains errors that show the model arrived at the same incorrect outputs is no proof. So, maybe this is at least a strong indication that it may have arrived at those answers by memorizing its training set at least to some extent. Unless one is fairly certain that particular forms of generalization are not needed—i.e., the model will simply not encounter scenarios not covered directly by the training data frequently enough for it to matter—it may be wise to get an estimate of the extent to which high scores are driven by data contamination. It may seem unlikely a model has encountered a particular input before, but the internet is big, and searching it is hard.
• How much overlap between a training and evaluation example should there be for it to count as ‘contaminated’? There are no easy solutions to these problems; approaches could include burning parts of common evaluation… However, to even get close to start addressing this insane issue requires access to training data, which are in many cases not accessible to people outside of the ones that trained the models in question.

Responsible LLMs

Even completely ignoring the potential consequences for humans and society, mistakes uttered by LLMs can have a direct impact on the viability of the product as well as the company itself.

Despite the importance of this topic, broadly evaluating accuracy, fairness, toxicity, factuality, and related topics remains incredibly difficult, both from a practical and philosophical perspective. The stakes of addressing it properly, however, are high, both for society and for model deployers directly: neglecting to properly evaluate these kinds of things can have a lot of inadvertent consequences for regulations, potentially making it impossible for the company to deploy similar models in their products.

Questions

• How far can automatic evals get us?
• Human evals are expensive.
• Are there ways to automatically evaluate higher-level skills?
• Data contamination
  • How should that be measured?
  • What is the impact of memorization on benchmarks?

(1) Open LLM Leaderboard or HELM

Maintainer: Hugging Face & Stanford respectively

Summary:
A sortable set of models, open and closed, evaluated on mostly academic benchmarks. Covers core scenarios such as Q&A, MMLU, MATH, GSM8K, LegalBench, MedQA, WMT 2014 along with other benchmarks. Largely open-source evals that have been developed over the years by various researchers. It encourages leaderboard hacking and faster saturation of academic benchmarks and wasn’t well thought through.

Developer goal:
General AI developers looking for models to build on will come here first, understand per a given size which models perform the best, and then leverage the result for their research, an application, fine-tune for a custom application. One thing that’s clear, however, is that there is a stage of disillusionment that the community is in when it comes to actual performance of models and how each rank on academic evals. This has driven the development of other leaderboards such as Chatbot Arena (see later).

(2) Hallucinations Leaderboard

Maintainer: Hugging Face

Summary:
It evaluates the propensity for hallucination in LLMs across a diverse array of tasks, including Closed-book Open-domain QA, Summarization, Reading Comprehension, Instruction Following, Fact-Checking, Hallucination Detection, and Self-Consistency. The evaluation encompasses a wide range of datasets offering a comprehensive assessment of each model’s performance in generating accurate and contextually relevant content.

Developer goal:
Have a starting point to understand which models hallucinate the least. There is some overlap with other available leaderboards here and, given without RAG or search integrated, these models will always hallucinate; this leaderboard is less interesting for developers.

(3) Chatbot Arena / LMSys

Maintainer: Together AI, UC-Berkeley, Stanford, and several others.

Summary:
A benchmark platform for large language models (LLMs) that features anonymous, randomized battles in a crowdsourced manner. Reliable in telling you which model is better, but can’t get into the details of specific capabilities (e.g., Gemini vs. GPT-4, where Gemini had internet access and caused confusion).

Developer goal:
Understand how models will perform in a real environment that challenges in ways that standard datasets/prompts can’t. Like #1 above, this may influence the developer’s starting point for further work.

(4) The MTEB leaderboard

Maintainer: Hugging Face & Cohere

Summary:
MTEB or “Massive Text Embedding Benchmark” contains 8 embedding tasks covering a total of 58 datasets and 112 languages. Through the benchmarking of 33 models on MTEB, establish the most comprehensive benchmark of text embeddings to date.

Developer goal:
Understand which models could perform best in applications where retrieval augmented generation (RAG) is used.

(5) Artificial Analysis

Maintainer: Independent / Startup

Summary:
Artificial Analysis provides benchmarks & information to support developers, customers, researchers, and other users of AI models to make informed decisions in choosing which AI model to use for a given task, and which hosting provider to use to access the model.

Developer goal:
Understand cost, quality, and speed trade-offs across models and model providers.

(6) Martian’s Provider Leaderboard

Maintainer: Martian (startup)

Summary:
Martian’s provider leaderboard collects metrics daily and tracks them over time to evaluate the performance of LLM inference providers on common LLMs.

Developer goal:
Understand quickly cost vs. rate limits vs. throughput for a variety of LLM providers and optimize for his or her application.

(7) Enterprise Scenarios Leaderboard

Maintainer: Patronus, Hugging Face

Summary:
Enterprise Scenarios leaderboard evaluates the performance of language models on real-world enterprise use cases (at present 6 benchmarks - Finance, Legal, etc.):

• FinanceBench (accuracy not retrieval)
• Legal Confidentiality (reason over legal cases)
• Writing Prompts (engagingness)
• Customer Support Dialogue (relevance)
• Toxic Prompts (generation)
• Enterprise PII (business-sensitive info generation)

Developer goal:
This leaderboard is still very much nascent as you can see from the lack of public models competing. Overall, though, this set of evals would provide developers with an idea of which base models to use as a starting point for a given task.

SANDBOX FOR DIFFERENT EVALUATORS AND SEE WHICH PERFORMANCE WHERE… SORT OF META EVALUATOR BY EVALUATING THE EVALUATION EVALUATORS

https://artificialanalysis.ai/

Unrelated to Patronus, but mainly Hebbia:

Knowing the prompt sensitivity of models could be valuable information for model deployers. Furthermore, differences in the way that prompts are selected make it impossible to compare performance for different models: is model A better than model B because it is better at the task, or did the developers of model A simply consider more prompts in their search?

Whatever prompting scenario is used, it is extremely important that prompt-selection efforts as well as prompts themselves are well-documented. Not doing so is not only harmful for reproducibility—in the process wasting the time of countless scientists that need to back-engineer the prompts with which reported scores were obtained—but also limits researchers to better understand what models are capable of.

Evaluating Outcomes

Similar to prompt engineering, using extracted, calibrated probabilities rather than generated text may be more informative about the upper-bound of what models can do than about what a model would actually do in practice, and it is interesting to consider the implications of that for the different stakeholders. For reasons analogous to the ones presented for prompt engineering, using calibrated log-probs may be an apt evaluation method for performance researchers: it gives a quick indication of what kind of knowledge the model actually contains and therefore constitutes an indicator of progress. For productionisers, however, it may be more appropriate to focus on generated output. While several preprocessing options may be conceived to address the variability that occurs as a consequence of prompting, the fact that a model has the correct answer somewhere in its probabilities is not particularly useful for a model deployer if there is no way to extract that from the model without knowing the answer in advance. As such, comparing the probabilities of answer A, B, and C instead of checking if a model actually generates A, B, or C may further distance the evaluation scenario and the scenario in which a model will be deployed. For an AGI researcher, using log-prob evaluations will evidently not suffice either: if a model in a multiple-choice task assigns a higher probability to option A than to option B, but it generates the letter D, an AGI researcher would likely argue that the model did not even really understand the task it had to do. Similarly, in many cases, normalizing probabilities may counteract the very idea by which in particular NLI tasks were constructed: predicting the relationship between two sentences independently of how likely the sentences themselves are is a crucial part of such tasks. In sum, focusing evaluation directly on generated model output might thus be a more suitable option, even if that results in lower scores.

Also from the Gemini 1.5 paper, the request for multimodal evaluations that perform well for long context evaluation was interesting.

Patronus Current evals fail with: Existing evaluations are increasingly strained by the new and rapidly advancing capabilities of large multimodal models. They typically focus on individual modalities and/or are restricted to tasks with shorter context lengths. Hence, there is a growing need for benchmarks which exemplify the nuanced requirements of real-world long mixed-modality use cases.

Qualitative long-context multimodal evaluations: Manually probe and stress-test the model’s long-context abilities, especially for novel capabilities where no quantitative benchmarks exist.

Especially relevant with their long-context evaluation call to action:

Thus, given the limitations of existing benchmarks and the challenges of human annotation, there is a pressing need for innovative evaluation methodologies. These methodologies should be able to effectively assess model performance on very long-context tasks while minimizing the burden of human labeling. To begin addressing some of these concerns, we recommend researchers and practitioners adopt a “multiple needles-in-haystack” setup for diagnostic evaluations which we observed to be signal-rich and more challenging compared to its “single needle” counterpart. We believe there is room for new benchmark tasks based on new or improved automatic metrics that require complex reasoning over long inputs (both human and model generated). This is also an intriguing research direction for creating challenging evaluations that stress more than just the retrieval capabilities of long-context models. We will be continuing the development of such benchmarks for realistic and comprehensive evaluation of model capabilities in the multimodal space. By addressing these open challenges and developing new benchmarks and evaluation methodologies, we can drive progress in the field of very long-context AI models and unlock their full potential.

https://rimba.cloud/

Send follow up mike wang intro to watershed employees

Schedule mike mona ESG guy

Do manufacturers want to take this process in house?

Larger firms are more likely to attain certification to an environmental sustainability standard - 37 percent of companies with over 100 workers had such a certificate, while only 5 percent of firms with fewer than 5 workers had attained such a status.

Rimba is compliance software for manufacturers to streamline certification reporting for their  annual audits. Their focus is on certifications like ISCC that focus on certifying sustainability across the supply chain. ISCC – as an example for this type of certificate – requires:

1. Traceability of sustainable products through the supply chain from the origin of the raw material to the final product.
2. Carbon emissions from the production of raw materials to the distribution of the final product.
3. Social and governance requirements suppliers must meet.

Broadly, compliance is a huge and growing issue in international trade. The requirements are numerous and vague which makes it difficult to implement and build a reporting system. 5 years ago, certifications were “nice to have’’. Today, regulatory pressures are making certification mandatory:

• EU = mandatory Renewable Energy Directive (RED) II since 2021
• Japan = mandatory through Japan’s revised Feed-in-Tariff (FIT) since 2021
• USA = mandatory through Renewable Fuel Standard (RFS) since 2005

Their target market is companies who deal with feedstock in their supply chain, thus needing ISCC and RSPO certification. Their initial traction has been with biofuel manufacturers in Southeast Asia, but they are also having conversations with potential customers in the US and EU. Examples:

• Apical (biofuel trader & largest existing customer) needs ISCC to sell to Italy.
• Valero (petroleum manufacturing, $55b market cap) needs ISCC Corsia to sell sustainable aviation fuel. CORSIA, the Carbon Offsetting and Reduction Scheme for International Aviation, is an emission mitigation instrument to address the increase in total CO2 emissions from aviation. 104 states (including all EU Member States), representing around 80% of international aviation, participate in CORSIA. Hence, the majority of airlines will be affected by the emission mitigation scheme.

ISCC and RSPO are the two top trusted certifications focused on biofuels and feedstock. RSPO is focused on palm oil and its derivatives, while ISCC has a broader certification scope for various feedstock.

TAM bottom up calculations

Knowns:

• RSPO has certified about 20% of the world’s crude palm oil (CPO).
• ISCC has certified a total of 57,000 companies.

Assumptions:

• ISCC and RSPO have certified a similar % (so 20%) of their respective markets, as ISCC covers other feedstock (rice straw, wheat straw, rice husk, and corn stover, used cooking oil, tallow, animal fats, etc).

Number of companies globally that should be ISCC-certified: 285,000 = 57,000 * (100/20)

Vanta pricing: For a company with a headcount of 200, the contract price ranges from $16,100 to $23,400. Companies with headcounts of 1,000 and over may expect to pay in the range of $20,700 to $92,200.

Assuming an annual contract value of $25K to $75K, it represents a market opportunity of $7.1 to $21.4 billion for Rimba to address.

Team

Distinct advantages to exploiting the Indonesian geography: Ryan was a Sr. auditor for certification frameworks like ISCC, RSPO, and RBO, and has built a network of renewable fuel suppliers and traders in Southeast Asia. Indonesia also represents 59% of the world’s supply of palm oil.

It is currently 2 co-founders who previously worked together in this industry, and have signed their first five customers (all small to mid sized biofuel manufacturers). The founding team is not set. They are dating a third technical cofounder responsible for building the software: Akshay Sharma (who’s already building for them), previously an ML engineer at Klarity, a Series A AI documentation platform. They applied to Y Combinator this Monday, and I believe there may be an opportunity for us to invest before YC if they get accepted.

What does their product do?

Manufacturers have to go through thousands of docs (PDFs) for each transaction every week. Docs can be (1) certificates of weight (weights of raw materials they receive), (2) contracts with their suppliers, (3) delivery notes, (4) sustainability declarations from certified suppliers, (5) purchase orders, (5) certificate of analysis, (6) invoices – to name a few.

These manufacturers already track all of these for their manufacturing reports. The docs are in paper format, so they can scan them, and info is used for general ledger. If companies are not certified, they only care about the weights of each shipment, but if they are certified, they need to start tracking a lot more information such as:

• Traceability of raw materials
  • Where did they come from? Factory’s location? Overlap with deforestation areas? Labor practices?
• Track real gas emissions
  • Calculate distance, type of trucks, how long of a journey, how much weight on trucks do they carry, depending on type of truck, type of fuel they use

Docs today are manually read, scanned, and added to multiple spreadsheets that track the aforementioned stuff. With all this data, companies create a mass balance spreadsheet that tracks all their inputs, and whether they are certified or not. This dictates how much % of their sales can go into markets that require certifications.

For example, if only 30% of their input materials are certified (in blended values) because they only manage to get 30% of their supply chain as certified. Then, only 30% of their sales can go into those markets that require certification (e.g., Europe and Japan). They can cross this limit, but it’s a loan to future certified materials, so in aggregate, they can’t cross this limit. If they do, their certification gets suspended.

The mass balance helps track this, how much certified material that comes in can’t be more than the certified biofuel coming out.

Things Rimba is building to help streamline this:

1. Data Extraction and Reporting Automation:
   • Process unstructured data e.g. scanned PDF, purchase order, weight certificate, shipping documents.
   • Automates data extraction for compliance reporting; Intelligent Document Processing (IDP)
2. Material Tracking and Risk Management:
   • Trace material flows and segregate certified from non-certified materials.
   • Prevents double accounting and maintains integrity of supply chains.
3. Streamline Multiple Sustainability Certification Audit:
   • more transparency + less manual data entry + less manual calculation on excel
   • Automates cross-certification, reducing effort and duplication.

Competitive landscape

There are a number of successful companies that help demonstrate that there is a market here including Watershed, Greenly and Sustain Life.

Watershed is a software platform to help companies to zero carbon. Their mission is to build the tools that enable a business to:

• Measure their carbon emissions to know where they’re at today.
• Plan out different ways to reduce the carbon emissions.
• Execute on that plan by buying clean power, by engaging their suppliers, by redesigning their product.
• Actually remove carbon from the atmosphere for whatever emissions remain (i.e., via their marketplace).
• Report on that progress externally so that investors, regulators, employees, customers, can hold them accountable (i.e., automate compliance for CSRD reporting) .

Podcast with CEO of Watershed (Scott Amyx);

Good news: There is starting to be some convergence and people are beginning to be in the space page about the rules of the road of what you count and how when tracking GHG. Bad news: The whole point of measurement is to enable the right action, and the status quo in carbon accounting maybe enables you to publish a really detailed PDF that has all the right acronyms on it, but does not enable companies to make the right choices. As an example, for most companies, most of their carbon emissions come from Scope3. The standard approach to calculating those emissions is to look at the dollars spent on a specific category and multiply it by what the EPA said in 2010 is the carbon per dollar in that category. What Watershed wants companies to actually do is redesign their products, change materials and change suppliers, and the status quo doesn’t enable that (i.e., just helps track but no actions). So Watershed wants to get the data to upgrade the quality of carbon accounting so it enables the right actions. So anyone can know the carbon footprint of a specific company, of a specific supplier, of a specific product, and you can make choices based on that.

In terms of regulations, the trend is for what’s currently being enacted in the UK and Europe to become standard practice around the world. Large companies, both public and private, are expected to report on carbon accounting at least once a year with the same rigor with which they report their financial data.

This category of carbon accounting software has gained crazy momentum in the last few years. Notably, both Watershed and Greenly were founded in 2019 and have announced growth rounds in the last 60 days:

• Watershed raised a $100M Series C led by Greenoaks, Kleiner Perkins, and Sequoia. Taylor Francis is cofounder/CEO.
• Greenly raised a $52M Series B led by Fidelity. Alexis Normand is cofounder/CEO.

The POV of Watershed is that if companies have pressure to take responsibility for their entire carbon supply chain, they can bring capital, resources, and expertise to decarbonize some of these lesser developed countries, because they are part of that company’s carbon supply chain. For example, when you look at a company like Apple, most of its carbon is coming from China and Taiwan from electronic suppliers (even though it appears on the US balance sheet).

This to me doesn’t connect with stats like RSPO having certified about 20% of the world’s crude palm oil (CPO). These producers of feedstocks like palm oil are based in the less developed countries that Scott (CEO) refers to in his podcast. 20% of crude palm oil having been certified already looks like incredible progress to me. Could it be that developed countries have this misconception that the biggest GHG emissions exist in less developed countries, when that might not be the case?

—

CSRD (Corporate Sustainability Reporting Directive) is another type of certification that applies to large EU companies and public and private non-EU companies with EU business. It replaces the EU’s legacy ESG reporting program—the NFRD—and raises the bar for breadth and robustness in sustainability reporting, covering categories beyond just carbon, including pollution, water, waste, and biodiversity.

They target companies in the EU and the US who are mostly concerned about carbon emissions (doesn’t matter if they are manufacturing, banks, or software companies). They are concerned about the mandatory disclosure they have to make about carbon emissions (i.e., CSRD). This is why companies like Greenly and Watershed have taken off.

Watershed also lets companies easily repurpose their CSRD data to fulfill other reporting frameworks, like CDP or ISSB. As an aside, only people who care about these reports are investors focused on ESG.

Watershed’s vision is to become the one-stop shop for corporate sustainability teams. They offer everything these teams need. It’s different to an approach like Rimba’s that’s going deeper into the end-to-end compliance requirements for a specific vertical (manufacturers) versus targeting a specific team (corporate sustainability) across many verticals. This shows in their product suite:

• Corporate: reporting; e.g., CSRD.
• Financial services: data to analyze portfolio’s emissions (ESG stuff)
• Supply chain visibility: gives corporations visibility over which suppliers are generating most carbon emissions (scope3 emissions i.e., GSG coming from suppliers indirectly), so they can grill them down and try cutting down their own carbon footprint.
• Consumer goods: help track carbon emissions for consumer good brands that care about consumer sentiment (i.e., make more money if they can show they are sustainable).
• Carbon Offset Marketplace: companies can purchase carbon offsets commensurate with their unavoidable emissions. The market for carbon offsets is expected to grow from $2 billion in 2020 to around $250 billion by 2050.

Watershed claims to streamline emission measurement using customer data to estimate a company’s Scope 1 (owned assets emissions), Scope 2 (utility purchases emissions), and Scope 3 emissions (3rd party emissions). For example, Watershed will use an Excel spreadsheet or a purchase order PDF to estimate emissions. Watershed also works with a company’s suppliers to accurately assess a company’s Scope 3 emissions. In November 2022, Watershed announced a supply-chain emissions tracking module to better address Scope 3 emissions with a “combination of public and private data from suppliers to estimate the emissions footprint of the larger supply chain.”

A dated view of Watershed’s product that is no longer on its website shows an expense item-level drilldown to illustrate the exact emissions coming from particular expense areas (e.g. advertising, software, etc.).

https://lh7-us.googleusercontent.com/__q4vhOyn42WZhJGOI55NWPqWmNCFLvvxjcCuvfc5BVYPLMWuxG7DDRc1Fqrvx8rVYXQuwdgiEvVM5XSRD9zsIgTIPFPGU4Vyu6dabamFkbTR7O7f3fAx6VDa1AWmb9McNMmLfSCr_vvPsF7KUpxzT8

Watershed’s takeaway: The majority of Watershed’s customers are technology companies. If climate reporting becomes mandatory, the market size will increase dramatically as corporations scramble to become compliant. You could potentially compare Watershed to ERP platforms like Oracle or SAP, or vertically-focused ERPs, like Veeva or Toast. If Watershed is successful the company would be considered a core system of record for managing carbon accounting. These types of ERPs can trade anywhere from 2-11x LTM revenue. Though in this podcast, Watershed CEO does describe their focus beyond carbon accounting, “we’re not in this to play the carbon accounting game, we’re in this to be a platform that enables decarbonization and companies at the front of climate change to make different choices”.

Why is Rimba different?

Rimba is focused on a different niche. They noticed that aside from carbon emissions, regulators are also starting to ask for disclosure of where the raw materials are coming from, social and governance aspects of these suppliers, and carbon emissions from the transportation that is involved in the supply chain until the consumer (upstream data). These data are captured in other certifications like ISCC.

Watershed and Greenly haven’t achieved any meaningful penetration more upstream in the supply chain (producers, traders, farmers). Most of their clients are startups, financial institutions, and public companies. When Rimba talks to potential customers like Valero (petroleum manufacturing, $55b market cap), IAG (International Airlines Group), or other manufacturers, they never mention Watershed as someone that is helping them.

Where Rimba comes in is that all these companies fail to understand how to navigate legacy industries like agriculture, manufacturing, and mining, which are all growing massively in Asia and LatAm supplying these products to the EU and the US. Most companies outside of the US and the EU probably never heard of CSRD certification or carbon emissions accounting. All they care about is how they can export products overseas, and for that they need to be certified with other kinds of certificates like ISCC, RSPO, and RBO.

Well, this still begs the questions.

1. Is it the logical next step for companies like Watershed to start adding other certificates to their platform after they win over CSRD compliance? If yes, which certificates are they likely to go after next? If the market opportunity for certificates like ISCC is obviously large, what are the reasons for why they wouldn’t choose to focus on that? Are there key differences between these markets? Is it a wholly separate problem/focus?
2. Why would Watershed’s lack of experience in legacy markets affect their chances of winning? What if the only important segment that needs to be won are large corporations serving the end consumers? One could say that if you win over this segment, you can help them force similar measures across all their suppliers upstream.

Per their CSRD website, Watershed’s likely to focus on the certifications next:

https://lh7-us.googleusercontent.com/YmWC33vb3B3-hLDtXWwMgbz6ODa7avoJ71LHPDVg_dtV03Z18dVXjzNNK-eD9iHDj5D5fYiSYyx3MajqxPSk15tS43staK2NyKsgrC3x2NhvfplB16LqNFSjdggxE8Z49YjQB_EbPrP-1wGimMZsRIM

Interesting potential customer use case:

Upstream data is unstructured (many suppliers are not sophisticated even in the US). Tim talked to the sustainability officer of a company called Highland Pellets and they sell biomass to the EU from the US. Highland sources wood chips and shingles from lumber mills and privately owned forests in the US. Many lumber mills and forest owners do not want to disclose anything. It is a constant push and pull, but Highland must be certified by certification: Sustainable Biomass Program / ISCC, and they need to track every square foot of where the trees come from. The requirements a few years ago were not very stringent, but it has gotten harder.

• If it is a certified supplier they would need less verification of supplier data. They would focus on delivery notes, sustainability certificate, and sustainability declaration.
• If the supplier is not yet certified, the company needs to verify the data themselves. (Like doing internal audit of the supplier)

Highland’s an interesting example as it shows how manufacturers struggle to get the necessary data from their suppliers. In Highland’s case the requirement is to trace where each tree comes from in which part of the forest. So Rimba has been discussing a solution by building a phone app for suppliers to take pictures of the trees before they are cut down, and Rimba will connect the app to use the photo and google maps to pinpoint the location. Right now Highland doesn’t have a solution for this yet.

CDP/SDDB/CSRD’s main audience is investors, these frameworks are meant to help public market investors determine which companies are better from a carbon accounting perspective.

ISCC on the other hand is more focused on operational data, and has a direct impact on a company’s ability to operate and do business. In order, it focuses on:

1. Traceability of raw materials
2. Deforestation, and then,
3. Carbon emissions

CarbonChain is a competitor that’s more similar to Rimba: Founded in 2019, CarbonChain is a carbon accounting platform specifically targeted at manufacturers, commodity traders, and their banks. Companies tied to physical supply chains have a more complex emissions footprint, and can also access unique sustainable financing if they can demonstrate emissions measurement and reduction. Customers include Lloyd’s, Concord, and Gunvor.

CarbonChain Comply offers a comprehensive solution for businesses within metals and energy supply chains who are grappling with monumental regulatory shifts on carbon emissions and rising stakeholder demand for transparent and accurate disclosure. The new one-stop shop for carbon reporting empowers sustainability leaders, compliance managers and procurement directors to streamline and satisfy every key disclosure request, no matter the format or framework. It raised a $10M Series A led by Union Square Ventures on April 10, 2024.

They are starting to talk about this other EU requirement called CBAM targeted at importers of aluminum, steel, iron, fertilizers, electrical energy, hydrogen or cement into the EU.

Interesting case study: https://www.carbonchain.com/case-study/concord

—

Detailed description of ISCC requirements (pdf p19 to p24):

The ISCC certification system covers three categories of requirements:

Sustainability requirements for agricultural and forest biomass, waste and residues, feedstocks for the production of renewable fuels of nonbiological origin and recycled carbon fuels and low iLUC risk feedstocks

Requirements for traceability and the chain of custody

Requirements for greenhouse gas emission savings and the calculation methodology

Sustainability requirements

Farms and plantations that produce sustainable agricultural biomass must comply with the sustainability requirements as stated in ISCC EU System Documents 202-1 “Agricultural Biomass ISCC Principle 1” and 202-2 “Agricultural Biomass ISCC Principles 2-6” ”. The six ISCC Principles cover the following topics:

Principle 1: Protection of land with high biodiversity value or high carbon stock

Principle 2: Environmentally responsible production to protect soil, water and air

Principle 3: Safe working conditions

Principle 4: Compliance with human and labour rights and responsible community relations

Principle 5: Compliance with land rights, laws and international treaties

Principle 6: Good management practices and continuous improvement

If biofuels, bioliquids and biomass fuels are produced from forest biomass the risk of using forest biomass from unsustainable production has to be minimised, and the biomass has to met certain land-use, land-use change and forestry (LULUCF) criteria. The use of forest biomass from sustainable production shall be ensured through monitoring and enforcement systems on national level or through requirements for the management system of the harvesting operation (Art. 29 (6) of the RED II). The six ISCC Principles cover the following topics:

Principle 1: Sustainability requirements for the production of forest biomass

Principle 2: Environmentally responsible production to protect soil, water and air

Principle 3: Safe working conditions

Principle 4: Compliance with human and labor rights and responsible community relations

Principle 5: Compliance with land rights, laws and international treaties

Principle 6: Good management practices and continuous improvement

Traceability and Chain of Custody

Traceability and chain of custody ensure that two basic requirements are fulfilled:

1. The possibility of tracing sustainable products back and forth through the supply chain from the origin of the raw material to the final product
2. The possibility of assigning product specific information to consignments (batches) of sustainable materials and products

Traceability describes the ability to identiy and trace the origin, processing history and distribution of materials and products through supply chains. Chain of custody describes the monitoring of input and output of sustainable materials and related information and documentation requirements. The following properties of sustainable material (so-called sustainability characteristcs) are the minimum information required under ISCC:

Type of raw material (e.g. rapeseed, sunflower, used cooking oil, etc.)

Country of origin of the raw material

Certification scope of raw material (e.g. the raw material is either certified according to the sustainability criteria of Art. 29 (2) – (7) of the RED II, or the raw material meets the RED II definition of waste or residue)

Information on GHG emissions (mandatory under ISCC EU, voluntary under ISCC PLUS)

Claim “ISCC compliant” or “EU RED compliant” (if applicable)

Greenhouse Gas Emissions

ISCC provides a methodology for calculating greenhouse gas (GHG) emissions for all elements of the supply chain and for determining greenhouse gas emissions savings. This can be applied to supply chains in all markets. Specific requirements apply for biofuels, bioliquids, biomass fuels, recycled carbon fuels and renewable liquid and gaseous transport fuels of nonbiological origin that are brought into the markets of the European Union. The economic operators bringing sustainable fuels into the market (i.e. the fuel supplier) must prove that the fuels where produced sustainably and that greenhouse gas emissions savings are sufficient according to the RED II.

The requirements for GHG emissions apply to all relevant supply chain elements from the production of raw materials to the distribution of the final product, including cultivation, collection and conversion processes, as well as the transport and distribution of intermediate and final products. Three different options are available to provide information on GHG emissions:

1. Use of total default values: Default values are raw material and process specific and are provided in the RED II for different types of biofuels, bioliquids and biomass fuels.
2. Use of disaggregated default values. So-called disaggregated default values are available in the RED II for the cultivation/ production of biomass, processing, and transport and distribution. Disaggregated default values allow the use of a combination of default and actual values.
3. To determine GHG emissions from cultivation the use of typical values that represent the average value in a specific area is also possible. Those typical values are often referred to as NUTS210 values. Those NUTS 2 or typical values have to be reported to the European Commission by Member States or third countries and can be used if the Commission recognizes them to be accurate . Use of actual values: Individually calculated values must be calculated based on the methodology according to the RED II.

All requirements to apply, calculate and verify the greenhouse gas emissions and emission savings are specified in ISCC EU System Document 205 “Greenhouse Gas Emissions”.

Participants in the Certification System

All economic operators that handle sustainable material (e.g. produce or generate, collect, process, store or trade) have to be covered by an ISCC certification. In the ISCC system the term ‘economic operator’ refers to a specific site (spatial entity) of a company where sustainable material is handled. A company may have more than one operating sites. If this is the case, every single operating site handling sustainable material has to be covered by its own certification.

For biofuels, bioliquids and biomass fuels produced from agricultural or forest biomass (including agricultural, aquaculture, fisheries and forestry residues) the supply chain starts at the farm/plantation or forest management unit (see Figure 2). The first gathering point is the first element that must be individually certified. Farms/plantations and forest management units can be covered under the certificate of the first gathering point but may also receive an individual or group certification.

https://lh7-us.googleusercontent.com/fTf_0Q63AkpeI3S4MJ2ij7C0oAFq6PKEeo0XqJvHh5oWqN6Mm6OBhBzlRF7_o7j3jnbAMbH1FjIzlYyPt51H00wMHDAenzKA92kHpHP5ceRjDSpUiPC6RW2r1HQXZFjx7VQK2t1OrIekqkg6FTfB0Uc

Biofuels, bioliquids and biomass fuels can also be produced from bio-based waste and processing residues. Recycled carbon fuels are produced from non-renewable liquid or solid waste. For renewable liquid and gaseous transport fuels of non-biological origin (RFNBO) it is the process energy that provides the energy input for the final fuel (e.g. electricity generated from wind, solar, aerothermal, geothermal or water). This process energy has to be derived from renewable sources other than biomass.

In these cases, the supply chain starts at the point of origin (see Figure 3). The collecting point is the first element that must be individually certified. Points of origin can be covered under the certificate of the collecting point but may also receive an individual or group certification.

https://lh7-us.googleusercontent.com/YYTV08v_9ykfbIAd5mtKnphUGrCk5TpXSG13YGeap7Or2SCxWM3Fp8AIKhk6Vbd3HE-YoY5Ux-gTAOb7THUG6iiYb7Ybm5OTeiBGE24kC75ySaz8YHm9c5_3tgZz0tv5pLBRNFyQFicjdXVviIwOBXQ

For all elements of the supply chain after the first gathering point or collecting point the certification requirements are the same regardless of the kind of raw material that enters the supply chain.

A valid ISCC certification is a prerequisite for any element of the supply chain that will handle sustainable material. The only exception to do this is for first gathering points and collecting points that may receive sustainable material up to three months prior to the start of the certificate’s validity. A detailed description of all relevant supply chain elements, the respective certification requirements and the requirements for handling sustainable material are stated in ISCC EU System Document 203 “Traceability and Chain of Custody”.

More random notes all over the place below

Their current workflow: Drive folder with all the pdfs per type, spreadsheet per type, mass balance report

Rimba is compliance software for manufacturers to streamline certification reporting for annual audits. They’re developing an AI enabled platform that uses LLMs and OCR to automatically extract key information from unstructured data sources (e.g. PDFs, scanned delivery notes, etc.). This addresses the challenge of manually handling thousands of such documents weekly, reducing human error leading to the risk of certification suspension.

The platform ensures accurate tracing of raw material origins and GHG (greenhouse gas emissions) calculation from transportations involved, which is crucial for meeting certification requirements. Based on customer feedback, this saves their customers from having to hire consultants, more employees to maintain certification, and reduce risk of suspension.

Their long-term goal is to expand into certifications in other industries such as textiles, plastics, food. They’ll first expand their product capabilities to integrate with various ERP softwares and – applying AI to cut down on manual review and calculations currently being done by consultants.

By accessing all data that currently lives on PDFs, they’ll focus on the following key features:

• Dashboard for visibility
• Trace material flows and segregate certified from non-certified materials.
• Prevents double accounting and maintains integrity of supply chains.
• Algorithm to match volume and source of purchased materials and sold materials
• Automates multiple certifications, reducing effort and duplication.

Rimba is compliance software for manufacturers. They streamline and automate how manufacturers follow mandatory sustainability certifications like ISCC, RSPO, RSB, starting with the renewable fuel market.

The market opportunity has grown and became more relevant in the last few years. The reason is that these sustainability certifications have become mandatory and require very stringent guidelines. Companies that don’t become certified can’t do business in the US, Europe, and Japan. The process for becoming certified and keeping the certification via annual audits lives in spreadsheets, manual work, and only happens thanks to expensive consultants. Needless to say, this is an expensive, time consuming, manual process. New technologies like OCR and LLMs allow for useful automations that can truly make a dent here, so there was never better timing to build a software company that helps automate and manage compliance for this market.

It is currently 2 co-founders who previously worked together in this industry, and have signed their first five customers (all small to mid sized biofuel manufacturers). The founding team is not set. They are still looking for a third technical cofounder responsible for building the software. They are actively dating CTOs and seem advanced with this one engineer called Akshay Sharma (who’s already building for them), previously an ML engineer at Klarity, an AI documentation platform. They are applying to Y Combinator by Monday, and I believe there may be an opportunity to invest before YC if they get accepted.

What does their product do?

Manufacturers have to go through thousands of docs (PDFs) every week. Docs are either (1) certificates of weight (weights of raw materials they receive), (2) contracts with their suppliers, and (3) delivery notes.

These manufacturers already track all of these for their manufacturing reports. The docs are in paper format, so they can scan them, and info is used for general ledger. If companies are not certified, they only care about the weights of each shipment, but if they are certified, they need to start tracking a lot more information such as:

• Traceability of raw materials
  • Where did they come from? Factory’s location? Overlap with deforestation areas? Labor practices?
• Track real gas emissions
  • Calculate distance, type of trucks, how long of a journey, how much weight on trucks do they carry, depending on type of truck, type of fuel they use

Docs today are manually read, scanned, and added to multiple spreadsheets that track the aforementioned stuff. With all this data, companies create a mass balance spreadsheet that tracks all their inputs, and whether they are certified or not. This dictates how much % of their sales can go into markets that require certifications.

For example, if only 30% of their input materials are certified (in blended values) because they only manage to get 30% of their supply chain as certified. Then, only 30% of their sales can go into those markets that require certification (e.g., Europe and Japan). They can cross this limit, but it’s a loan to future certified materials, so in aggregate, they can’t cross this limit. If they do, their certification gets suspended.

The mass balance helps track this, how much certified material that comes in can’t be more than the certified biofuel coming out.

Things Rimba is building to help streamline this:

1. Data Extraction and Reporting Automation:
   • Process unstructured data e.g. scanned PDF, purchase order, weight certificate, shipping documents.
   • Automates data extraction for compliance reporting; Intelligent Document Processing (IDP)
2. Material Tracking and Risk Management:
   • Trace material flows and segregate certified from non-certified materials.
   • Prevents double accounting and maintains integrity of supply chains.
3. Streamline Multiple Sustainability Certification Audit:
   • more transparency + less manual data entry + less manual calculation on excel
   • Automates cross-certification, reducing effort and duplication.

Some context on how the market works:

Independent auditors run third party verification services that audit whether manufacturers are following certification guidelines.

Consultants help train and prepare manufacturers for these audits. In practice, manufacturers only learn 20-30% of what consultants tell them in Year 1, so they continue paying for their services annually to make sure they don’t get suspended in the annual audits.

Manufacturers also hire an in-house team (typically under the legal department) that make sure all the required data is gathered.

ISCC and RSPO are focused on biofuels and feedstock. 50K companies have been certified, a 25% of those that need it, so the total number of companies here is 180-200K.

Knowns:

• Some certifications are bullshit (self reported/voluntary guidelines), but others are real (third party verification via regular audits) and have real economic consequences (not allowed to sell in various markets).
• Acquiring and maintaining these real certifications is not an easy task. It’s manual, time consuming, and expensive. Currently, consultancy services like Control Union are the biggest players in helping manufacturers navigate the certification requirements (5000 employees, $600M annual revenues).

Unknowns:

• Unclear how much money manufacturers spend on consultancy services + in house employees to navigate certification processes.
• TIC companies are involved with this, but unclear how relevant is certification for the largest companies in this space.

Testing, Inspection, and Certification (TIC) market.

• Certification is part of testing, inspection and certification and the provision by an independent body of written assurance (a certificate) that the product, service or system in question meets specific requirements.It is the formal attestation or confirmation of certain characteristics of an object, person, or organization. This confirmation is often, but not always, provided by some form of external review, education, assessment, or audit.

The Global Testing, Inspection, And Certification (TIC) Market is projected to grow at a CAGR of 5.5% from 2022 to 2030. the market was valued at USD 233.69 Billion in 2021 and is expected to reach USD 389.79 Billion by 2030.

• SGS S.A. ($7B annual revenues, public)
• Bureau Veritas S.A. ($6B annual revenues, public)
• Intertek Group Plc. ($3.5B annual revenues, public)

Testing, Inspection, And Certification (TIC) Market industry is experiencing unprecedented growth, driven by regulatory compliance, outsourcing trends, environmental awareness, and technological advancements

• Rising Regulatory Compliance: Governments across the world are enforcing stringent regulations to ensure the safety and quality of consumer products, compelling manufacturers to engage independent TIC services.
• Outsourcing Trend: Manufacturers are increasingly outsourcing TIC services to specialized providers, aiming to meet compliance standards while streamlining operations and enhancing profitability.
• Environmental Awareness: Growing environmental consciousness has led to a surge in environmental testing and certification, driven by both governmental mandates and consumer demand.
• Technological Advancements: Continued innovation in testing, inspection, and certification technologies will enhance the industry’s efficiency and capacity to address emerging challenges.

Another relevant market: revenue attributed to environmental certificates is expected to increase from $37.8 billion in 2023 to $102.3 billion in 2032, at a compound annual growth rate (CAGR) of 11.7%.

Bottom up TAM analysis on SOC2 compliance annual spend:

https://start.askwonder.com/insights/market-size-compliance-audit-software-uoqrw344c

Therefore, we can assume that the total estimated SOC 2 spend in the U.S. every year would be (75%9,000)60,000 =405,000,000.

• Total number of SaaS companies: 9000 [USA]
• Avg annual audit spend: $20k to $100K
• 75% of SaaS companies get SOC certification

Similar numbers for ISCC alone:

• Total number of ISCC certified manufacturers: 50,000 [Global]
• Avg annual audit spend: unknown; Tim’s mining company $20K.

Customers

Focused on the Indonesian market for palm oil; Indonesia produces 44.5 million metric tons of palm oil, or roughly 59 percent of the world’s supply. It’s in so many everyday products around the world—shampoo, deodorant, cereal, dishwashing detergent—that much of the world could scarcely function without it. Yet the massive industry is still lacking in many basic regulations. The high-yield nature of palm oil, which yields almost ten times as much product per hectare as soybean oil, makes it a lucrative prospect for corporations. And its expected market growth could nearly double to £87 billion (over $105 billion in US dollars) by 2026—what Bintang’s team calls “an agricultural gold rush.”

Largest customer: https://www.apicalgroup.com/sustainability/progress-dashboard/

• https://rspo.org/members/2-0235-11-000-00/

Klarity

A disproportionate percentage of highly skilled, highly paid workers waste time reading similar documents over and over again.

Klarity’s platform was built over the course of four years and leverages cutting-edge Natural Language Processing and Computer Vision to understand documents as well — or better — than human reviewers. The result: The most accurate document intelligence platform available today.

Market size

Environmental certificates is a growing market. Seemingly because the carbon and energy attribute certificates can be traded in carbon markets, and a lot of money can be made that way. It’s still unclear to me how third-party verified certificates like ISCC for biofuels or Energy Star for energy efficiencies are connected to this. My guess is that if I am a biofuel producer, and now I get 40% of my production to be certified materials, perhaps I can create credentials from that that I can later sell in the markets?

Unclear the market size breakdown. How much of it is related to financial markets stuff like carbon bonds, certificates that are sold or bought by other companies to offset carbon footprint versus the actual cost of managing certification compliance. I assume that number is much smaller.

Different kinds of certifications:

• Self reported or voluntary guidelines
• Third party verification and mandatory

like ISCC, RSPO, and RSB in the context of BioFuels

His plan:

• Technical cofounder
• 10 companies under their belt

They have 5. Get another 5.

RIMBA is a compliance automation platform for manufacturing companies globally to meet industry-standard sustainability certifications.

I am looking to understand what is everything that needs to be tracked. Why not just narrow down their focus to ISCC to start? Each certification seems to be a pretty unique process (different to what he somewhat suggested in the first call that they are all kind of the same). I want to understand in detail how can software truly move the needle for ISCC (specifically one example).

1. Extracting data from invoices.
2. Who does traceability? Are companies good at this with or without software? Startups seemed to tried doing traceability but not from compliance angle. Thoughts?
3. Who tracks traceability, and who tracks greenhouse gass emission savings? How do companies get to these numbers? How do they create their methodology?

Isn’t $200k a bit too low for this to be a big problem?

ISCC requirements:

https://lh7-us.googleusercontent.com/ko9nf–aMmXk3GUZ_CrAKkkvr2ska_I6XY3PYzl2n3nQ3GrdmH_9ovANaN22Hu5QYTBn3iI5vKrJgKbgs8uDjmtnMxcUV7ia0DDpoSm0thMEGM_ycddMm4ySQxCLohYyKKpu3I0FVLNR6E4lGlNFQZ8

Product traceability in supply chain is the process of tracking a product’s origin and journey through the supply chain, from manufacturing to end-use. It involves logging manufacturing information in a database and following products at every step.

https://lh7-us.googleusercontent.com/FawjajLnI1UBkn2LJFgWqrWSTbEI-JIOzsHRG62qV42JjGK_jAON4C6WbxOF9tdWvjq9c7k4tSPcNj_JnVXsqL-AIktAI3wwUrNB1yYFQJnNcbEs0XGSuZj9sRizlzQDFEVfHGqX080zhJoduTWjkas

Lack of product expertise. As it is right now, I worry they don’t have the experience or talent to build a great modern product to serve how this industry manages compliance. They claim to be Vanta for this sustainability certification niche, but I worry in practice they won’t get close to building a similar quality product to Vanta’s. However, I do think there is a good founder market fit with the CEO and this space, and if somehow paired with the right CTO, this could become a killer team.

• Beyond the idea. We want to see a first-gen product that’s delivering value.
• Product-centric opportunities, e.g. fundamentally superior product.
• PMF focus. Startup with a maniacal focus on validating product-market fit.
• Builders. The product is a core competency of the team, not an afterthought.
• True entrepreneurs. Not in it for the immediate gratification, glory, lifestyle.
• Deep insight. Founders understand the problem better than most people in the world.
• Blue ocean strategy. Product creates new market space rather than competing head-to-head.
• Wedgeable. Nearterm path to dominating an initial, narrow wedge of a larger market.
• Interesting IP created by solving very hard problems worth solving.
• Venture scale. The business should be raising VC to capitalize on its opportunities.
• Scalable business model. If not initially, then very possible in the intermediate term.
• Durable business model. The company can survive/thrive outside of trends or economic cycles.
• Execution. There is evidence that the team can do the hard work.
• Pre-A. We will work with founders to get them into a nice series A.

Market

• Does this match our stated areas of interest?
  • If not, does the opportunity represent a new potential area of interest that we should further research?

It has the potential to be interesting. Staying compliant with many of these certifications looks like a cumbersome, manual process where software has very little penetration. If built the right way, the opportunity of building a Vanta for this type of compliance seems appealing to me.

• Do we intuitively understand and agree with the problem?

Obtaining and managing a sustainability certification is a manual, time consuming process. If it costs $200K a year to manage compliance (auditors, consultants, skilled labor, and certification) for a $20M revenue company, it looks like a relatively high cost.

• Do we agree with the company’s perspective on the market?
• Is this a blue ocean or red ocean market?
• How big is the market, both immediate and potential?

No numbers so far but relatively large. There are a lot of manufacturers that need to be certified, and for that have to manage their business.

• What are examples of successful companies that demonstrate that there is a market here?

There are no successful large scale product startups in this space.

This one seems to have failed: https://jobs.techstars.com/companies/bioledger. They were trying to build a single source of truth for all transfers in the lifecycle of a biofuel. Not sure what they were trying to do or who this was for.

Product traceability in supply chain is the process of tracking a product’s origin and journey through the supply chain, from manufacturing to end-use. It involves logging manufacturing information in a database and following products at every step.

https://www.veriflux.io/. This other competitor also emphasizes the traceability aspect of this. RIMBA didn’t discuss this as much. Why?

• If not based in the US, are there distinct advantages to exploit in that geography?
• Is there a wedge approach to the market?
• Is the business model scalable to the degree that warrants venture capital?
• Is the business model durable enough to survive surprising dips in a market cycle or unexpected tailwinds?

Team (Founders)

1. Empathy: This entails a deep understanding of the customer’s needs and the broader market pain points. It’s not about superficial knowledge, but rather a profound comprehension that fuels the team’s instincts and drives their decision-making process. This empathy allows them to connect with their customer base on a more intimate level, creating products and services that truly resonate.

2. Expertise: This isn’t just about having domain experience—it’s about possessing a level of understanding so deep that it sets the team apart from others. Their expertise should enable them to navigate complex industry challenges and uncover unique insights that can be turned into actionable strategies.

3. Execution: This is all about the nitty-gritty of operations. The team’s ability to execute plans efficiently and resourcefully is key. It’s not just about having a great idea—it’s about being able to bring that idea to life in the most effective way possible.

4. Experimental: This quality is characterized by a willingness to build, test, learn, and iterate. It’s about having an open mindset and the flexibility to adapt and evolve. A team with an experimental approach is unafraid of making mistakes—they see them as learning opportunities that can lead to improvement and innovation.

• Are the founders good at communicating the problem?
  • Did they teach us anything new?
• Do they understand this problem better than most people in the world?
  • What gives that insight or experience?
• Is this the right team to capitalize on the opportunity? Even if they have the right insight, make sure that this is the team to do it.
• Are the founders hands-on builders?
  • What have they built in the past?
• Who is responsible for difficult, technical work on the team?
• Would we describe this team as true entrepreneurs?
  • True entrepreneurs are resourceful and will survive the low points and exhausting grind of building a business. They are not in it for the immediate gratification, glory, lifestyle, or treat it as a stepping stone in their careers.

Product

• Is the company working on directly addressing the “problem,” or are they working adjacent to it?
• Is it clear who is the intended audience for this product?
• Is the product built and launched?
• What makes the product fundamentally superior or different from other options?
• Does the product deliver value to users/customers in its current state?
  • What are the top 2 reasons that users/customers get value from the product?
  • What are the top 2 improvements that users/customers would benefit from?
• Is there interesting technology or IP being built here?
• What are the associated costs to this product? Consider the cost of building and distributing the product.

Traction (Momentum)

• Do the founders believe they have validated a minimal degree of PMF?
• How do the founders view the process for figuring out PMF?
• How has this team shown that they can execute on their potential?
• What are recent inflection points that make this a good time to invest in the company?
• What are the primary drivers to their growth?

Opportunity

• Once they raise money, what is unlocked for the company moving forward? How does this accelerate them?
• Who are the Series A investors that we think would be interested in this company, if they hit the right milestones?

What big questions have been answered so far? Derisk. Validation/invalidation

Before solutions like Vanta, obtaining SOC2 certification was generally more manual and time-consuming. Here’s a breakdown of how it typically went:

• Understanding the Requirements: Companies had to deeply understand the SOC2 criteria, which involve detailed controls on security, availability, processing integrity, confidentiality, and privacy. This meant a lot of research and possibly hiring consultants to ensure full comprehension.
• Manual Documentation: Businesses needed to manually document all processes and controls relevant to the SOC2 standards. This often required extensive efforts from internal teams or expensive external consultants to map out processes, write policies, and prepare the required documentation.
• Implementing Controls: Implementing the necessary controls without a guiding software meant relying on internal knowledge and often bespoke solutions. This step could include configuring systems to meet security standards, establishing new protocols for data handling, and training employees on compliance-related issues.
• Evidence Gathering: Companies had to manually collect evidence proving the controls were in place and effective. This was often a labor-intensive process involving screenshots, logs, and records which needed to be meticulously organized and maintained.
• Auditor Interaction: Interactions with auditors were typically more hands-on and continuous, as companies had to demonstrate their compliance through meetings, emails, and document submissions without centralized platforms that streamline these interactions today.
• Remediation: Any gaps found during the audit process had to be addressed manually. This often meant revisiting and revising processes, which could be time-consuming.
• Continuous Compliance: Maintaining compliance was an ongoing effort without automated monitoring tools. Companies had to regularly check their controls manually to ensure they remained compliant over time.

https://lh7-us.googleusercontent.com/3s0XBn201Pop5bzlrrUf6bKYgbnfPB7oothV4oZ3yDTRDpGehPLiieSEXSSPxDfJKq5XSQE9Ll9IBdEvjisnMmyq3z6ind5sVrRk6Q8BJtgzg0dCfcq8-QW7Ueii9kwQjseXGHZD3vMldmPW0KGzuUc

SOC2 has technical requirements that need to be constantly monitored such as MFA authentication. What are some examples of things that need to be constantly monitored in the context of ISCC? What are the things that need to be shown / managed?

• MFA not on all resources, Vanta is integrated with all ur systems and identifies when that isn’t the case.

https://lh7-us.googleusercontent.com/44z6HLPJwerBePnuowiireA7xRSgmcbaBriOxA3YzvdP4AfFrkuvk-odzbcaWMJ6stH83O3RQ1vqEfTvk4mDQcp-A6DqXpqi0GW2OBUitWMunlh2HuzLptDIKMYHSLrEsxw13APb8NXU8ISmsm2Vw1E

Going /against https://www.ibm.com/products/openpages

Yc application:

I dont like company description. “Made easy” what does it mean? What does the company do?

Why turbotax ? why not sth like Vanta

Rimba is a software company to help manage sustainability certification compliance for manufacturers, starting with the renewable fuel market

Carbon carbon accounting software has gained crazy momentum in the last few years. Notably, both Watershed and Greenly have announced growth rounds in the last 60 days:

• Watershed raised a $100M Series C led by Greenoaks, Kleiner Perkins, and Sequoia. Company started in 2019, and Taylor Francis is cofounder/CEO.
• Greenly raised a $52M Series B led by Fidelity. Company started in 2019, and Alexis Normand is cofounder/CEO.

Greenly:

https://lh7-us.googleusercontent.com/VUnhQ7cq1D3xdLqQ7QFtp7iG6BJwB-YNA5LXuniKNzd1UpzYfxuSJRWGMbmTmsJsnTN15SoFvlyBe5XxT50fjPIWc7T48g_oyxe4DsfcwN7fX8N4zyW9B9cLUEIz7yrDIyPQFKY4g2-P06s20CK76WY

Gadi: “Plenty of major corporations have announced net-zero emissions targets. And while that alone won’t be enough to move the needle on climate change, these targets have gotten the ball rolling. Those net-zero goals are trickling down, too, influencing companies throughout the supply chain.”

Naturally, all of this trickles down into the manufacturers in markets like Asia and LatAm too, bcs if corporates care about CSRD, assuming that generates even more pressure for suppliers to care about other certifications like ISCC

Tim: Correct, companies in EU and US are mostly concerned about carbon emissions (doesn’t matter if they are manufacturing, bank or a software company). They are concerned about the mandatory disclosure they have to make about carbon emissions. That’s why companies like Greenly, sustain.life, watershed have taken-off.

A niche we noticed is that aside from carbon emissions, regulators are also starting to ask disclosure of where the raw material is coming from, social and governance aspect of these suppliers, and carbon emissions from the transportation that is involved in the supply chain until the consumer. These are captured in certifications like ISCC. This has not really taken off yet, so we still have a head start.

Gadi: I saw greenly helps corporates push certification measures to their suppliers. If this works, they could slowly onboard the complete supply chain into their platform to help trace carbon footprint. I’d be curious if they found some success with this

The question Id ask myself too is: Is it the logical next step for companies like Watershed to start adding other certificates to their platform after they win CSRD? If yes, which certificates are they likely to go after? If the market opportunity for certificates like ISCC is obviously large, what are the reasons for why they’d choose not to get into that market? Are there some key differences between these markets that would make it such that if Watershed expands, they’d start losing focus because this is a different problem/market?

Tim: Both watershed and greenly are more focused as a carbon accounting platform

https://lh7-us.googleusercontent.com/YmWC33vb3B3-hLDtXWwMgbz6ODa7avoJ71LHPDVg_dtV03Z18dVXjzNNK-eD9iHDj5D5fYiSYyx3MajqxPSk15tS43staK2NyKsgrC3x2NhvfplB16LqNFSjdggxE8Z49YjQB_EbPrP-1wGimMZsRIM

Gadi: Greenly says they automate and streamline the data collection process, but they don’t show how their product works. Do you know how they do it?

Expertise: I am still struggling with this. The founders have clear domain expertise, and understand the process of ISCC extremely well. However, they fell short in painting the complete picture for his market. Perhaps initially due to my bad questions, but I later discovered Watershed/Greenly as two companies that are going after the carbon accounting software and a few certifications related to this by the EU. I had previously asked Tim for more successful competitors in the space, as I was alarmed his only references were dormant startups, so him not sharing this initially worried me.

• Does he think Watershed/Greenly are not direct competitors of Rimba? Was he trying to not mention these because they are already too large?
• Is it the logical next step for companies like Watershed to start adding other certificates to their platform after they win CSRD? If yes, which certificates are they likely to go after? If the market opportunity for certificates like ISCC is obviously large, what are the reasons for why they’d choose not to get into that market? Are there some key differences between the market Rimba is targeting versus Watershed’s such that if they expand to it, they’d start losing focus or be at a major disadvantage against Rimba?

CONVOS WITH CHATGPT

The data requirements for ISCC (International Sustainability & Carbon Certification) tend to be more granular and specific to the sustainability of biomass and bioenergy products, focusing deeply on the environmental impacts and traceability of raw materials. In contrast, the CSRD (Corporate Sustainability Reporting Directive) covers a broader range of sustainability data, encompassing environmental, social, and governance (ESG) factors applicable across all operations of a company. Here are examples of the types of data required for each:

ISCC

1. Source and Origin of Biomass:
   • Type of Data: Specific details about the biomass source, including geographical location, type of biomass (e.g., agricultural residue, specific crops, waste materials), and the method of cultivation or collection.
   • Purpose: To ensure the biomass is produced or sourced sustainably, without contributing to deforestation or biodiversity loss, and to verify compliance with legal and sustainability criteria.
2. Greenhouse Gas (GHG) Emissions:
   • Type of Data: Detailed calculations of GHG emissions throughout the lifecycle of the bio-based product, from cultivation and harvesting of biomass to processing and transportation.
   • Purpose: To document reductions in GHG emissions compared to fossil fuel benchmarks, supporting claims of environmental benefit and compliance with sustainability goals.
3. Sustainability Practices:
   • Type of Data: Information on agricultural practices (e.g., use of pesticides, water management, soil conservation practices), labor conditions, and compliance with social standards.
   • Purpose: To ensure that the production of biomass adheres to environmental and social standards, promoting sustainable development and safeguarding worker rights.

CSRD

1. Environmental Impact:
   • Type of Data: Broad data on energy use, water consumption, waste generation, recycling rates, and direct and indirect greenhouse gas emissions across the company.
   • Purpose: To assess the company’s environmental footprint and identify areas for improvement in resource efficiency and environmental stewardship.
2. Social and Employee Matters:
   • Type of Data: Information on employee diversity (e.g., gender, age, ethnicity), labor practices, health and safety statistics, and human rights considerations within the company and its supply chain.
   • Purpose: To promote social equity and worker welfare, ensuring that the company’s practices align with societal values and labor laws.
3. Governance Practices:
   • Type of Data: Details on the company’s governance structures and practices, such as board composition, executive compensation, anti-corruption and bribery policies, and shareholder rights.
   • Purpose: To provide transparency about how the company is managed and governed, ensuring accountability and alignment with the interests of stakeholders.

The ISCC’s requirements are more focused on the specific environmental and social impacts of biomass production and use, requiring detailed tracking and reporting to ensure sustainability throughout the product lifecycle. In contrast, the CSRD requires a wider range of data, reflecting the company’s overall impact on sustainability, including governance aspects. Each type of reporting serves different purposes and targets different aspects of sustainability, with ISCC being more narrowly focused and CSRD encompassing a broader corporate perspective.

Understand workflow

What the product does

What does their product do?

Manufacturers have to go through thousands of docs (PDFs) every week. Docs are either (1) certificates of weight (weights of raw materials they receive), (2) contracts with their suppliers, and (3) delivery notes.

These manufacturers already track all of these for their manufacturing reports. The docs are in paper format, so they can scan them, and info is used for general ledger. If companies are not certified, they only care about the weights of each shipment, but if they are certified, they need to start tracking a lot more information such as:

• Traceability of raw materials
  • Where did they come from? Factory’s location? Overlap with deforestation areas? Labor practices?
• Track real gas emissions
  • Calculate distance, type of trucks, how long of a journey, how much weight on trucks do they carry, depending on type of truck, type of fuel they use

Docs today are manually read, scanned, and added to multiple spreadsheets that track the aforementioned stuff. With all this data, companies create a mass balance spreadsheet that tracks all their inputs, and whether they are certified or not. This dictates how much % of their sales can go into markets that require certifications.

For example, if only 30% of their input materials are certified (in blended values) because they only manage to get 30% of their supply chain as certified. Then, only 30% of their sales can go into those markets that require certification (e.g., Europe and Japan). They can cross this limit, but it’s a loan to future certified materials, so in aggregate, they can’t cross this limit. If they do, their certification gets suspended.

The mass balance helps track this, how much certified material that comes in can’t be more than the certified biofuel coming out.

Things Rimba is building to help streamline this:

1. Data Extraction and Reporting Automation:
   • Process unstructured data e.g. scanned PDF, purchase order, weight certificate, shipping documents.
   • Automates data extraction for compliance reporting; Intelligent Document Processing (IDP)
2. Material Tracking and Risk Management:
   • Trace material flows and segregate certified from non-certified materials.
   • Prevents double accounting and maintains integrity of supply chains.
3. Streamline Multiple Sustainability Certification Audit:
   • more transparency + less manual data entry + less manual calculation on excel
   • Automates cross-certification, reducing effort and duplication.

Further breakdown of requirements:

What makes them special?

How do they convince people to work this way?

Rimba is focused on X, while Watershed is focused on Y. But it does help demonstrate that there is a market here because of Z.

Below, I’ll explore the opportunity and indirect tax compliance market:

Sales tax is very hard.

Sales tax is hard because the tax code is complex, unstandardized, and decentralized across 14,000+ jurisdictions in the US alone. Governments make businesses figure out how to comply with this complexity. Businesses are responsible for knowing where, when, and how much tax to collect; maintaining state registrations; and filing returns.

This is not just true in the US, it’s a global problem.

There are new startups focused on solving it for other geographies like Europe (i.e., Taxdoo) and Latam (i.e., Brinta), while incumbents like Avalara are looking to expand their presence internationally.

Intro to Sales Tax.

This morning, when you bought a coffee from your favorite local cafe, you probably paid 6-10% more than the sticker price because of this pervasive and ubiquitous thing called sales tax. Sales tax is embedded in virtually all commerce in every country, state, and municipality around the world. It’s a major source of revenue for state, local, and national governments.

https://lh7-us.googleusercontent.com/MPlR16O6GmRRIGPu2FhrS2Rjxa-X3Sg0n_5St99sj-oNL0KV-JYNAl0r3gLNiha1oudbA3cf2jI-jO7CS2PoeZBf6kR1FbzGI5bWW67KJus-ORTdn6AijHa2i4Vnl91ufy7CYLyl3FNIpdM9eIpsOBY

https://lh7-us.googleusercontent.com/TqUWs99-1t9tqoRKfyDBMXLsL0j3bFbbj4SFQLOm9HyZCZQSiugELaZAusJ_IRkFV_95UuKfoVGJCeMd4anbwIltXsqW1gOMtZIvObcSWTGKzHYzoJ9Lwgly0DSQHcOM7AZ7w-B9poP–202bhjsd8s

https://lh7-us.googleusercontent.com/nEPD7G9r7FxV3q6EQ7fQ3mjprU4yhUcle4lmXdFj81AfsxycbRhPRxHNdx3sAmPVTHVQSfAKTuSPta5RFyX-Ms433rx3FwzA76i2HMk6JDmN7nyju_tGgztmrzpOeZG96r7oNeTA7qTOLY0QySAcNQQ

The ancient Egyptians were the first civilization recorded to collect sales tax. Our modern sales tax system in the US traces its roots to the Great Depression when 25+ states established state-wide sales tax regimes to bolster revenue. Today, all but 5 states enforce sales tax. There is no escaping it if you consume or do business in our modern world. As Benjamin Franklin famously said:

“in this world nothing can be said to be certain, except death and taxes”.

https://lh7-us.googleusercontent.com/YnaQGRSFV1sUSOVJqbEDW2m7HwI8O8AQj1pcsWqRI-E9VR_FuyrpDc-ID9YNgbtNzp6a_qUtTjDNP0gsVk2v-ABsGGC1MjmGg2_7yRb1VeUKYjpYP3JnoRr5vk5sbXwE1IucFaRLGLIyo351DKIKEyw

Why is Sales Tax so complex:

• Sales tax law is regulated by individual states, counties, and cities. There’s no universal standard. The tax law and rates in Ohio are different than in California.
• The US has ~14k different taxing jurisdictions. In 2022, there were 6k+ tax law changes made within those jurisdictions. The rules are complex and can vary dramatically depending on the industry, products sold, where and how they’re sold, etc.
• The tax code was written for a world in which commerce was primarily conducted in person. The internet changed all that. Today, every business is default national or global from day one. Businesses of all sizes deal with sales tax on a scale that was previously faced by only the largest Fortune 500 enterprises.
• Historically, internet companies only had to collect sales tax in states where they had a physical presence, also known as “nexus” (retail stores, distribution centers, offices, etc.). Meaning most companies weren’t collecting sales tax in the majority of states where they sold products. A 2018 ruling called Wayfair vs South Dakota changed everything, allowing states to enforce sales tax compliance on online purchases regardless of whether or not the company had a physical presence in the state. Today, 45 states and the District of Columbia enforce sales tax on internet companies.
• Marketplace facilitator laws are another recent regulatory change. These laws force marketplaces like Amazon, Walmart, Etsy, etc. to collect and file sales tax on behalf of their merchants. States started passing these laws in 2019. Currently, 30+ states and counting have passed some form of marketplace facilitator law.
• The growth in remote work has also increased the burden of tax compliance because hiring a contractor or FTE in a new state triggers an immediate tax compliance obligation, even if the company has no sales in that state.
• Merchants don’t just sell physical products anymore. Now you can sell software, AI-generated art, e-books, data, subscription boxes, and whatever else entrepreneurs and creators make. Figuring out how your products are taxed is a challenge for modern companies that may not fit squarely into traditional boxes.

What customers expect

The last few years have seen tremendous innovation across B2B financial products. Today, you can set up payments (Stripe, Adyen); payroll (Gusto, Rippling, Justworks); banking (Mercury), corporate cards (Ramp, Brex, Expensify); hire globally (Deel, Remote), etc. — all in a matter of minutes. Hours at most. You don’t need to be an expert. You simply need to click and let modern software do its job.

https://lh7-us.googleusercontent.com/Anc741PzHseT-y-D7e5PxKaagB8guBsmD0NvFDbQi1lgQhcqpKy-p8zANDyc1VJOSdc68fgWlYLujHppUP9uQiXVD0PsvKXD9rt645qSoxIAsdjxlBxwbtZnPs4TVWfrTknLjEv4kKvAf29q6OBtZNA

Finance and accounting leaders have come to expect easy, intuitive, and elegant products, but that’s not what they get when shopping for a sales tax solution. The fact that getting sales tax compliant is still so difficult is what initially sparked most new tax startups’ curiosity.

Problems with existing software products

Today, you need to know a considerable amount about sales tax to successfully implement a sales tax product or pay someone who does. Avalara (the mid-market market leader) and Vertex (the market leader in the Fortune 500) sell complex enterprise tools meant to be used by in-house tax teams and/or set up by Big 4 accounting firms. Implementation can take months, hundreds of hours of engineering/accounting resources, and cost tens of thousands of dollars.

These are the most common frustrations…

• Implementation is time-consumptive, painful, and expensive
• Pricing is cost-prohibitive for growing companies
• UI is confusing
• Not developer friendly
• The product tax codes don’t make sense for SaaS companies
• Mapping your products to the tax code is manual and confusing
• Customer support is terrible
• Missed or late tax filings
• Taxes being overpaid
• Inaccurate reporting data

Customer research

Whatt I’ve seen from online reviews

Avalara fucked up our setup, which caused us to not collect the right amount of sales tax. The difference came out of our pocket.

CEO of an 8-figure ARR, Series A SaaS company

The product works if you get it set up right, but it just sucks. Their pricing is outrageous. They charge you for API calls that expire. So even if you don’t use them all, you lose them going into the next billing period.

CEO of a venture-backed alcohol marketplace

It would have made our business unprofitable to comply with sales tax using Avalara so we decided to just risk getting audited until we have more revenue.

CFO of a bootstrapped $7-figure digital products marketplace

From start to finish, it took a bunch of time to implement Avalara. Probably 8 weeks… 3 engineers, me, and my accountant.

CFO at a $9-figure+ e-commerce marketplace

Because one of our products is taxed as Telecomm, we had to use two different APIs because the Telecomm product is different than the AvaTax product. Which made our engineers lives miserable for a couple months. We’ve actually still not finished getting it setup.

Controller at a ~$40M ARR, Series B SaaS company.

I never know what I’m supposed to do with the information that Avalara gives me on a given state. Sales tax is risk management, and unless I pay them or an accounting firm 10k, I’m left shooting in the dark.

CFO of a 7-figure ARR venture-backed SaaS Company

We ended up with $1M in liabilities plus penalites from different states because we started growing too fast. Our controller didn’t know what to do. So we tried to go to Avalara to get it fixed, but they ghosted us on our first meeting then tried to charge us $3k for something called a nexus study before they could give us a quote….

Founder of an 8-figure e-commerce holding company

Stripe Tax was too expensive and incomprehensive. Avalara was also very expensive, but at a certain business scale they’re the only credible option…. We probably spent about 30-40 engineering hours on the Avalara implementation in total. Not including my time or my CFOs.

CEO of an 8-figure ARR Series A SaaS company

Our accounting firm told us to use Avalara…. They ended up registering us in too many states and then didn’t do the registrations right in some states. So I had to go figure out how to manually log-in to each state’s website and figure out what was going wrong…. Can’t believe I spend $50k a year on this.

CEO of an e-commerce marketplace

Implementation was a bitch. The docs were inaccurate. We had to get on multiple calls with their support engineers just to figure out what was going on…. The total setup ended up taking like 6 months.

Engineering Manager at $80M+ Series D SaaS Company

I’d rip out Avalara tomorrow if a better product came long

CTO of a $100M+ e-commerce brand

Avalara overpaid one state by $100k for one of my clients. We couldn’t even get ahold of their customer service to fix it.

Managing partner of a sales tax CPA / consulting firm

Market

This industry is the classic story of a dominant legacy incumbent starting to show important weaknesses that newer competitors are looking to exploit. Mainly, it is a red ocean market opportunity, with a go-to-market strategy that heavily relies on flawless execution.

The distinction, though, as early-stage investors, is that there are many startup competitors as well that have entered the market in the last 12 to 24 months.

What the market looks like today

Incumbents:

• Avalara: Comprehensive tax compliance solution with (almost) global coverage. Known for its strong branding but has faced customer frustrations. They were bought out by Vista for $8.4B in October, 2022.
• Vertex & Sovos: Enterprise-grade solutions with deep tax expertise. Sovos is more prominent in Europe. Vertex is a publicly traded company with a $3.5B market cap, and Sovos had been bought by Vista and more recently owned by HgCapital.
• TaxJar: User-friendly, e-commerce focused. Struggled with the SaaS market due to limited partnerships with billing platforms.
• Stripe: Attempted to dominate billing, payments and tax, but faced challenges in execution and integration post TaxJar acquisition.
• Thomson Reuters: a ~$60B public company that also has a suite of legacy enterprise tax products.

Startups:

• Anrok (most similar to Taxwire): Focused on post PMF SaaS companies. They have the most developed product and business for SaaS sales tax of any startup to date. They raised a $20M Series A in 2022 at a $100M valuation, led by Index Ventures and Sequoia Capital, and are rumored to have around $5M in revenue.
• Zamp: Primarily focused on e-commerce brands. They’ve chosen to take a very services-heavy approach to what they’re building.
• Numeral: A startup from the most recent Y Combinator batch. They’re focused on handling filing & remittance for e-commerce companies. Biggest strength is that the founders have e-commerce backgrounds.

First, enterprise-graded companies are unlikely to switch providers. Most of the very large companies already have a solution (and those that don’t may choose not because they do highly concentrated sales). The long term opportunity of targeting the younger SaaS market is to grow as many of these startups scale, but there’s no immediate opportunity today to win big accounts.

Second, Avalara absolutely dominates the mid-market. Here, it is Avalara versus everyone else. Their core product was designed in the 2000s. The rest of it has been stitched together with the 18+ acquisitions they’ve made since. Internally, their tax filings are entirely manual. The product experience is clunky and overwhelming for non-experts. Their tax codes aren’t configured for tech companies. Their international product is weak. For example, large enterprises like Mindbody (Avalara customer) are expanding internationally and choosing to go with alternative products for international tax compliance.

If you talk to Avalara’s customers and partners, you’ll discover that they’ve created a lot of frustration and animosity in the market. Since being acquired by Vista, customer service has declined, compliance mistakes increased, and pricing has gone up. Avalara’s NPS is 13.

TaxJar (founded in 2013) was the closest thing to a real Avalara alternative, but Stripe has de-prioritized it since acquiring the company in 2021 (rumored at ~$400M), laying off most of its team and halting GTM. Remaining efforts by Stripe in this space are focused on Stripe Tax.

Stripe Tax is not a true enterprise product. It can be good enough for small companies, but the fees, tax engine breadth, and lack of compliance features (filing, monitoring, registrations, etc.) prevent it from being a real solution for growing companies. Stripe Tax has access to Stripe data, and falls short in analyzing a company’s broader context.

Global sales tax compliance is a growing market. States and countries aren’t going to stop collecting taxes. Recent regulatory changes like Wayfair have increased compliance burdens. There are still many non-compliant companies that will need to get compliant. States are cracking down and the number of audits seem to be rising.

Concluding remarks

There are two main reasons why now may be good timing for a new player to enter the space: (1) market changes, and (2) increased regulations:

• Avalara has been acquired by PE, and their product isn’t built for what many customers need today. They’re also more focused on moving up-market (competing with Vertex for Fortune 500 clients), cutting costs, and raising prices. They’ve created a lot of frustration among their customers and partners, and TaxJar is no longer a major player, so there’s no clear challenger to Avalara.
• Regulatory changes like Wayfair vs South Dakota and marketplace facilitator laws coupled with increasing scrutiny from the states have made sales tax compliance a larger priority than ever.

However, there is no clear opportunity to leverage newer technologies (i.e., LLMs) to disrupt the industry. A few examples of suggested uses of LLMs:

• Ffacilitate a much faster onboarding process by understanding the business situation of customers using LLMs. During onboarding, users can explain their business in natural language, give a link they can scrape, so they can better automate the understanding and explaining of a company’s taxability.
• Building a tax engine from scratch involves aggregating and interpreting large amounts of text and converting that into math. LLMs have already proven powerful for us in moving faster on understanding and interpreting the tax code.

The opportunity, as I see it, is to build best-in-class modern software with a great user experience that every company loves using. I’ve struggled to find a product approach that would afford a startup an unfair/intrinsic advantage in its go-to market, which is what we’re looking for in software opportunities we back. Tax compliance is a low-margin-for-error problem space, offering no room for true innovation. At the end of the day, most startups can’t control the regulations and how each State treats taxes, and companies are forced to play by their rules.

A big challenge is that the pull for tax compliance arises at specific growth milestones, so it is not a problem for most younger companies until it is.

(Here is the spreadsheet I used for the graphing process - I gave access to will@contrarycap.com and web@willrobbins.org )

The venture-capital industry – originating in the US – is envied mainly throughout the world as a rear engine of innovation and economic growth. However, people with a distorted sense of globalization tend to underestimate the need for adapting to different markets while operating VC funds. On one side, popular media portrays the resurgence of protectionism and nationalism through an emphasis on deteriorating trade agreements, the US-China trade war, placement and removal of tariffs, and political backlash in countries with austerity programs. Meanwhile, the rapid rise of online connectivity – especially social media, has overshadowed reality to give an inflated view of a flattened world. The combination of both phenomena, thus, gives rise to a skewed understanding – “the world is very connected but decelerating due to socioeconomic and political tensions.”

The authors of the global connectedness index aim at correcting the distorted view by proving that globalization isn’t a one-time phenomenon but a process of change through time. They reframe globalization as a process using four metrics – trade, capital, information, and people – at different levels of analysis. The intensity of the relationships and variety of those relationships among countries.

Depth of the metrics in how much the country engages in each activity shows the intensity in international business compared to localized movements. Although, as VC money circulates through international boundaries, it’s affected by cyclical and structural economic determinants. Visible economic growth could be attributed to the structural factors – indicating increasing connectedness. However, volatility in cyclical economic factors and macroeconomic situations could cloud results from being completely attributed to globalization.

While the breadth of the metrics implies the range of regional relationships by the VC, breadth is profoundly affected by “The Law of Distance”, indicating interactions are dampened by distance along the sociocultural and geographic reach. In addition to “cultural affinities, political agreements, and infrastructure links,” as the startup ecosystem of a country develops along with viable market size, the prospect for investment increases. However, due to (1) the considerable investment risk that goes along with VC funding, and (2) slow growth curve for new investment hubs to develop; the breadth is limited as VCs become risk-averse towards new regions. Breadth remains relatively stable with slow growth.

Thus, the author’s usage of a cumulation between depth and breadth to indicate globalization indexes apply to VCs. The weight of the average might weigh more towards depth, as breadth might be significantly harder for VCs to penetrate. Even in an ideally connected world, in-depth insider knowledge of the economy and sectors within a country is needed to enable investing and expanding to newer hubs.

The four metrics are used to functionalize globalization by embodying critical indicators of a truly global market. The metrics provide cohesive rhetoric in defining globalization. Although each of the pillars affects and is affected by global expansion, some are more relevant depending on the industry. For venture capital and private equity firms, the movement of people, though significant, is of least relevance. Trade and information flow accompany VCs in the process of globalization. Unsurprisingly, the capital pillar has more sway in functionalizing VC globalization.

As markets become attractive, there are multiple ways to infuse money to high-growth, high-yield startups, and companies alike. Money gets injected through trade deals like bilateral treaties that favor taxation avoidance, where shares are exempted from taxation. Secondly, through FDI, given that legislative barriers permit such activity for VC and PE funds through a structure usually known as a Financial Foreign Direct Investment(FFDI). Third, through joint ventures in regions with lesser expertise and more legal hurdles.

The circulation of more capital reduces the tangible and intangible trade costs, creating trust, and thus generating a competitive advantage for local, innovative companies to flourish. This economic development must be distributed in emerging countries as well to boost globalization. Foreign funds face hurdles when investing in emerging markets. One hindrance is the lack of appropriate exit strategies that give sizable returns to angel and seed-stage investors. IPO stock markets or acquisition strategies are underdeveloped or non-existent. Even with the growth and expansion of local firms, a phenomenon of “semi-globalization” remains in developed countries.

!https://s3-us-west-2.amazonaws.com/secure.notion-static.com/e451beb5-b23d-4ba6-89b1-df7126cc628f/Untitled.png

Per the argument above, I used the cumulative Global Connectedness Score to analyze countries on the steepest growth and decile. Using the SLOPE function to indicate the steepness of points, Mozambique and Iran showed extreme values of the rise and drop respectively

Iran, a country with the 4th biggest oil reserve, occupies a critical geographical area in the middle east. The reform process that started in 2005 radicalized the Shia Muslim country to adopt a strict Sharia Law. Funding Hezbollah - a terrorist group against Israel and enrichment of Uranium caused regional tensions to rise and culminate in UN sanctions imposed in 2006 that resulted in economic blockade from the rest of the world. After sanctions, prices of goods dramatically rose and the startup scene got affected.

As the UN put sanctions on the country, outside investment and trade that was incoming shut off. Due to the rising living conditions within the state, startups, and independent businesses - that are premier targets for VCs - weaned off and growth declined. Second, the venture capitals that were highly concentrated around western powers, and China feared the political instability.

Classifying for general failures of venture capital investment in Iran, it could be due to formal and informal barriers. Significant formal deficiencies are similar to emerging markets. They include a lack of financial regulations, inefficacy in tax, labor, financial disclosure, investment protection laws, and capital market underdevelopment. Iran’s informal barriers widened the formal institutional voids. Strains include a culture that heavily disapproves of capitalism and an infiltrated culture of secrecy.

Although local firms operationalize and are making moves, they are usually government sponsored. This hamper and disturbs private VC funds into making biased and manipulated investment decisions. The displacement of funds from private investors leads to a lack of mature VCs that create a self-sustained venture capital market, that can’t even develop in Iran – let alone expand internationally.

Mozambique’s story is of rising from demise with the right reforms.6 The country had been underdeveloped due to the two-decade-long civil war. However, after 1994, the economy had been in recovery mode and reviving. Coincidentally, the natural resources of the country shot it into the global playing field as aluminum exports proliferated, followed by the discovery of vast reserves of natural gas and coal in the early 2000s.

International economic support with newly discovered resources diversified the country’s economy supporting old industries of agriculture and tourism. Throughout these years, the government of Mozambique set attractive legislative and policy practices that made the right balance of national development and attraction of foreign investment for building a long-term, inclusive, and sustainable ecosystem.7 Plus, the government introduced privatization of formerly state-owned entities, especially in agriculture and natural resource exploitation.

FDI, in the form of venture capital and private equity in Mozambique, is incentivized by the integration of the country to the Southern African Development Community (SADC) that opens access to member countries. Even with substantial constraints of administrative, infrastructural, and bureaucratic difficulties seen in Iran, the government’s reforms have significantly moved the country in the right direction. Given the reform-minded government, the attractiveness of the market remains with significant rewards that keep pushing Mozambique as a good investment location in Africa.

Fickle Pickle Bar is a neighborhood bar next to the residence hall that I lived in Seoul for 4 months. Tuesday nights, the rooftops right sit was reserved for my roommates and I.

I looked at this spot in HBC, Seoul, that I liked and thought of as an excellent investment to explore further. Fickle Pickle is a multipurpose complex located in the trendy and “hip” neighborhood of Seoul. Due to the massive influx of youth in the area, the owners – 2 Korean ex-NYU students – decided to found a space that could attain the ever more so “fickly” youth of Seoul. Space tries to incorporate a café, bar, rooftop, and art studio in one space. The studio features new, independent, up and coming artists to be featured every month. Due to the alternative, grunge, and industrial feel of the place, plus the artist featured, Fickle Pickle, is growing within this international district.

Post discussions with one of the managers, Jen, I constructed financial statements with justified assumptions set in place. The café, located on the ground floor, is open every day. While the studio, bar, and rooftop are open 4 days from Thursday-Sunday. Sales from the café, drinks from the bar and rooftop, entrance fee for the studio, and Fickle merchandise (including T-Shirts, pouches, stickers, key chains, and caps) are the primary sources of revenue for the complex. An hour-long observation suggested close to KRW 819,570,000,000 to be made in sales per year.1 Exhibit 1 below shows how that revenue is stripped to result in an 11.8% profit margin.

!https://s3-us-west-2.amazonaws.com/secure.notion-static.com/527d4617-cbca-4c33-8732-890ce7f8986a/Untitled.png

                   Exhibit 1: Income Statement for Fickle Pickle ‘19

COGS included fixed and variable costs, of which higher percentages were salaries and cost of the product/service offered. As Fickle Pickle’s customers are mostly people within the HBC community and featured artists get the utility of exposure from the venue, SG&A costs are relatively minimal. Overall profit margin is 11.83% which is quite high in relation to the concentration and margins of coffee shops and similar service businesses in South Korea.

Marketing, R&D, and Human Capital used here are expensed, which might not necessarily be the sole truth. As Fickle Pickle has additional intangible assets on the “vibe,” they create and the brand image they maintain that creates a feedback loop of attracting more customers and increasing the value of the business. So be aware of the distortion this might induce in underestimating the intangible assets.

!https://s3-us-west-2.amazonaws.com/secure.notion-static.com/f9f03a42-2b66-4b95-94af-6e1f90eb237a/Untitled.png

                    Exhibit 2: Cash Flow Statement for Fickle Pickle ’19.

The positive cash flow in 2019 entails that the business owns easily liquidated cash for operations and fulfilling short-term obligations and shows working capital’s health. This cash shows of sustaining cases of emergency expenses, demand for higher inventory, and ability to stabilize supplier power. Fickle Pickle is gradually trying to cut down on inventory at hand for better turnover rates. Having positive cash flow shows they have the necessary funds to get inventory in case of a higher demand to negotiate with supplier power to keep up with demands while minimizing surplus. As there is no debt financing cash flow and new equity isn’t issued, the total for financial activities is 0. Capital Expenditure increases as the complex get new material like a new beer tap and renovation expenses for the bathrooms.

!https://s3-us-west-2.amazonaws.com/secure.notion-static.com/3675f402-ce29-4902-8def-8b9c8630f1ed/Untitled.png

                     Exhibit 3: Balance Sheet for Fickle Pickle for ’18, ‘19

Considering Fickle Pickle is a well-established venue in the community with strong customer relationships, I used a moderate/high discount rate of 10 %. When using the DCF model and estimating the current value, I omitted the terminal value, heavily due to the fact Korean businesses have a short time lived and get replaced quite remarkable. Otherwise, buyouts by bigger conglomerates or the intense market competition cause business to dissolve. Note that when taking this approach, I undervalued a key input to the DCF model and have a smaller NPV, and you should not want that if you intend to keep your money there for a longer time, it generates profits. The NPV of the business ended up being approximately KRW 516,000,000, with the book value at 2019 being KRW 603,543,150.

During this evaluation, I assumed the lack of terminal value as businesses within Seoul are usually short lived and easily regenerated. Although the business might have enduring value, its not going to give value tremendous value long down the line big enough to need terminal value considerations. The NPV of the business ended up being approximately KRW 516,000,000 with the book value at 2019 being KRW 603,543,150.

As Fickle Pickle was a small business, raising debt would have imposed unnecessary financial burdens. However, now that it has its core audience base of loyal customers, raising additional debt wouldn’t leave a constraint as they have a consistent cash flow. Loss of company control isn’t required to raise additional capital as well. So. on the next phases of growth and expansion like opening another branch or significant capital expenditures, Fickle Pickle should borrow money as its revenues and margins show stabilization over the years.

Being placed in a competitive environment with premium pricing needs Fickle Pickle to be efficient and offer a distinctive utility and experience to its consumers. As seen in the assumptions, we demand a higher price from consumers, and boosting prices, might have damaging impacts leaving pricing as a sensitive lever we can’t pull. The service offered, however, has significant room for improvement. From increasing the number of days open (after customer maturation) when the cost of production and service for the new days(Mon-Wed) are exceeded by the net profit to be made.

!https://s3-us-west-2.amazonaws.com/secure.notion-static.com/896e1abc-397d-4c0d-b820-f099378138ee/Untitled.png

One of the most significant growths is SG&A expenses to leverage the incoming traffic at Seoul’s “hottest” district, increasing the visibility of their presence and value proposition in the market. Plus, maintaining the service quality, number of customers/hours is going to increase with improved SG&A expenses. There is a peak number on the number of consumers/hours, as the current location isn’t big enough. The reach of optimal instore numbers is an indicator of expansion to new locations to capitalize on demands.

Looking at these prospects for Fickle Pickle, and understanding the intricacy of the investment process, I advocate for investing in the company as it will provide return in coming years.