PROJECT TITLE

CoAI Corp: Understanding What Learning Algorithms Do With Data, Plus an AI Memory Platform to Fund the Research

SHORT SUMMARY

Machine learning research on how learning algorithms and data interact, paired with an AI memory product that funds the work. This microgrant covers Delaware incorporation, cloud compute and research tools.

FUNDING GOAL

Minimum: $3,000

Target: $5,000

PROJECT SUMMARY

We are a pre-incorporated research and technology startup doing machine learning research while building a commercial AI product. Everything so far has been self-funded. A few external contributors help voluntarily from time to time. They are not employees or paid contractors.

RESEARCH

We try to understand what learning algorithms actually do when they interact with data. We know how training works. The harder part is understanding the system that comes out of it, and why it behaves the way it does. Our process is Observe -> Question -> Hypothesize -> Reason -> Test -> Verify.

To show the kind of approach we follow, a good example is the paper "Exploiting Vocabulary Frequency Imbalance in Language Model Pre-training" by Woojin Chung and Jeonghoon Kim (NeurIPS 2025): https://arxiv.org/abs/2508.15390

Instead of accepting that "bigger vocabularies help," the authors ask where the benefit comes from. They scale vocabulary from 24K to 196K while holding data, compute and optimization fixed. They find that the gain comes almost entirely from lower uncertainty on the 2,500 most frequent words, and that the same gain appears when model size grows with a fixed vocabulary. Our research works the same way: we do not just ask whether something works, we ask what the system is actually doing and why.

PRODUCT

We are building AI memory and tools around existing frontier models, so a model does not start from nothing each time. A very good memory system is something we can build ourselves without training a large foundation model, so we put our work into the memory architecture itself. It includes:

- Static memory

- Optimized retrieval

- A custom retrieval system we built ourselves

- Multi-embedding models (embeddings from Google)

- A cache architecture

Some details cannot be disclosed because they are part of the product. People use frontier models together with our memory and tools. The product is rolling out soon at https://chat.coaicorp.com

We are not training a large model right now because it needs a large amount of compute. Any commercial model we produce later will be available on the same platform. The product exists to fund the research over time.

WHY THIS MATTERS FOR AI SAFETY

A trained system is not programmed like normal software, so we cannot just know what it is capable of. We have to investigate how the algorithm and the data behave together. Research that explains why trained systems behave as they do is a basis for knowing what they can do and for predicting how they will behave. That is what our research aims to build.

GOALS AND HOW WE WILL ACHIEVE THEM

1. Incorporate as Co-Artificial Intelligence Technologies Corp. (CoAI Corp.) in Delaware. Without a formal company it is very hard to raise money from investors, and many cloud providers and credit programs require one.

2. Run more experiments with proper compute. We will use cloud resources and credit programs to test our hypotheses on how learning algorithms and data interact.

3. Launch and improve the memory platform at https://chat.coaicorp.com and iterate on the memory architecture.

4. Publish openly. We will release our findings, papers, tests and code on GitHub and our research site.

This grant is an initial microgrant, not the full budget. It gets us the legal structure and compute access we need to apply for larger programs and raise investment.

HOW WILL THIS FUNDING BE USED?

We are asking for approximately $3,000-$5,000:

- Delaware incorporation: ~$500-$1,000

- Cloud computing and credits: ~$1,500-$2,500

- Research and software resources: ~$500-$750

- Small reserve: ~$500-$750

At the $3,000 minimum, incorporation and a core compute budget come first. Additional funds go mainly to cloud compute and research tools.

TEAM AND TRACK RECORD

[Add founder name(s), roles and short bios. Include relevant research experience, previous projects and publications.]

- A few external contributors help voluntarily from time to time. They are not employees or paid contractors.

- The company has been fully self-funded to date.

LINKS

Company: https://coaicorp.com

Research: https://research.coaicorp.com

Philosophy: https://coaicorp.com/philosophy

GitHub: https://github.com/COAICORP

LinkedIn: https://www.linkedin.com/company/coaicorp/

Product (rolling out soon): https://chat.coaicorp.com

WHAT ARE THE MOST LIKELY CAUSES AND OUTCOMES IF THIS PROJECT FAILS?

Several things could go wrong:

- A hypothesis could be wrong. Research does not always confirm what we expect.

- An experiment could show nothing. A null result is still informative, but it does not produce a headline finding.

- The product could fail to get adoption. If so, we would have less revenue to fund the research.

If the project fails, we will open-source our research papers and tests so others can learn from them. Even a failed project should leave something useful behind.

HOW MUCH MONEY HAVE YOU RAISED IN THE LAST 12 MONTHS, AND FROM WHERE?

None. Everything has been self-funded.