I want to tell you the truth from the first paragraph. The architecture is validated. The performance numbers are real — measured in a live environment running on my own trading account. The economics are calculated from empirically observed inference data. But the company is not yet registered. There are no paying customers. There is no team beyond me. The system is not commercially deployed.
I am writing this because I need investment to make those things real. And I believe the case for that investment is compelling precisely because of what one person was able to build without any of them.
This is a seed-stage investment opportunity in a fully validated but pre-production AI system. The architecture is proven. The funding takes it to market, registers the company, builds the team, and begins the scaling the architecture was built to achieve.
What I Built
Neural Alpha is a three-layer ensemble architecture for autonomous cryptocurrency portfolio management. Three specialist AI models, each optimised for one cognitive function and one only:
Layer 1 — SniperLSTM: 30 timesteps of market history across 512 features per asset. Bidirectional LSTM with 4-head attention. One question: what temporal patterns are forming? 5.52 million parameters.
Layer 2 — Transformer Encoder: 60-step market snapshot across 128 features. Six layers, 8-head attention. One question: what does the full market context say? 3.55 million parameters.
Layer 3 — RL-PPO Agent: Receives the assembled outputs of both models — 180 features per decision — and acts: what position, what size, which assets? 4.68 million parameters. The only layer that adapts continuously from live trading experience.
The principle behind the three layers came from watching how expert human traders actually work. A technical analyst, a macro analyst, and a portfolio manager are not three people doing the same job. They are three specialists each doing a fundamentally different cognitive task. The architecture instantiates that division in software — and the performance gain comes directly from it.
The entire system runs on a commodity CPU — no GPU, no cloud GPU, no specialised hardware. Total inference time per cycle: 1.479 seconds out of 19.3 seconds. 7.67% active computation. The rest is warm-idle: models loaded in memory, waiting for the next cycle. This is why the economics work.
The Discovery
When I designed the SaaS architecture to share this system with multiple users, something became clear that I had not anticipated — and that I cannot find described in any existing literature.
Every user has a different account balance, joins at a different time, has a different risk tolerance. Every user's agent therefore encounters the market from a different position. Every trade it makes deposits a distinct experience — what market conditions it saw, what action it took, what outcome resulted — into the shared RL training buffer. The RL agent trains on all of it.
A brain trained on a thousand users has seen the market through a thousand different lenses. It knows things no single-user system could learn, because the diversity of user experience covers regions of the market's state space that any individual would take years to encounter alone.
"More users produces a smarter brain. A smarter brain produces better returns. Better returns attract more users. The loop closes itself — at zero coordination cost between any two agents."
No agent communicates with any other. No coordination is designed. The intelligence accumulates in the shared training buffer — the same principle that makes the brain smarter than any neuron, that makes markets aggregate information no individual holds. I looked for this specific structure in the multi-agent AI literature. It is not described. The combination — independent execution, shared learning, zero designed coordination, recursive improvement from population diversity — is the architectural innovation.
The scaling law that follows is architecturally guaranteed: more diverse users, smarter shared brain, better performance for every user. The magnitude of that gain at 10,000 users has not been measured because the deployment does not yet have 10,000 users. Getting to that measurement is what this funding enables.
The Economics
Infrastructure cost — why the margins are structural
| Infrastructure basis | COGS / user / month | Status |
|---|---|---|
| GPU cloud — always active | $84.00 | Superseded |
| GPU cloud — warm-idle optimised | $0.177 | Superseded |
| Commodity CPU — compute only | $0.013 | Validated |
| Commodity CPU — full stack at 1,000 users | $0.064 | Validated |
Revenue projections — based on validated unit economics
| Scale | Monthly infra cost | MRR (projected) | Gross margin |
|---|---|---|---|
| 100 users | $6 | ~$3,900 | >99.8% |
| 1,000 users | $64 | ~$39,000 | >99.8% |
| 10,000 users | $175 | ~$394,000 | >99.9% |
| 100,000 users | $1,250 | ~$3,940,000 | >99.97% |
Infrastructure costs are validated. Revenue projections assume $39/month average revenue per user across the tier mix. What is not yet validated is user acquisition — which is what the distribution budget in this round funds.
The competitive moat is not the margin. The margin is a consequence of the moat. The moat is the scaling law: every user who joins makes the brain smarter for every other user. A larger node has encountered more of the market's state space. The intelligence differential between a large-population node and a small one widens automatically over time, without any investment in R&D — because it is driven by the structural operation of the training loop.
What the Investment Funds
Entity registration, regulatory compliance for algorithmic trading in target jurisdictions, IP protection, legal agreements. The infrastructure that makes commercial operation possible.
Exchange API integrations, production infrastructure, security audit, monitoring systems, user onboarding. Taking the validated architecture from a working system to a commercially deployed product.
Two to three ML engineers. The architecture is built and documented in 107 pages of technical specification. The team accelerates the multi-node recursive architecture and cross-domain extension.
The first 500–1,000 users. This is the seed population for the stigmergic training loop. The diversity of their experience is what begins building the intelligence differential that compounds automatically from that point.
Milestones this round funds: company registered, product commercially deployed, first 500 paying users, first empirical measurement of the intelligence scaling curve across different population sizes. The next round is justified by those measurements.
This Is Bigger Than Trading
The stigmergic training loop — diverse agents depositing experience in a shared medium, collective intelligence compounding without any designed coordination — is not a trading idea. It is the structure that intelligence takes whenever it appears in nature.
Neurons in the brain. B cells in the immune system. Ants building colonies. Humans building economies. The same four-component pattern at every scale, in every substrate. Neural Alpha is the first deliberate engineering instantiation of that structure at computational speed. The trading domain was the laboratory. The principle is general.
Five coupled nodes — molecular binding, cellular response, pharmacokinetics, toxicity, population efficacy — discovering drugs that work through the entire pipeline, not molecules that pass one stage and fail at another.
Every patient encounter deposits experience into a shared diagnostic brain. A hospital system builds collective clinical intelligence from the full diversity of its patient population — available to every clinician simultaneously from day one.
Atmosphere, ocean, land, biosphere, and human activity as coupled nodes — the coupling learned from real-world observations, not estimated from historical parameters that fail in novel regimes.
A system that learns from every monitoring agent in every country and distributes improved collective understanding back to every health authority simultaneously — not after months of data collection but continuously.
I am not asking investors to fund these applications now. I am asking them to recognise that the mechanism proven in the trading domain is the same mechanism that produces those results in other domains — and that the trading deployment is the proof of concept that validates the general principle. The company I want to build is the infrastructure for that principle. The trading engine is the first product, not the last.
What I Am Offering
I am open to multiple structures. The right partner matters more than the specific deal shape. What I am not willing to compromise is the research direction and the mission — this architecture exists to benefit humanity, not merely to optimise for returns. That is non-negotiable in any structure.
Standard seed investment in exchange for equity. Founder retains operational control and research direction. Investor gets board representation and pro-rata rights in future rounds.
Investment from a partner — a healthcare company, AI infrastructure firm, or research institution — who sees the cross-domain architecture and wants to co-develop the next application domain alongside trading.
Full acquisition by a company with the infrastructure and vision to take the architecture to its full potential. I remain as lead researcher and architecture director. The mission is preserved.
The Honest Assessment
I am a health science graduate who built this alone in Nairobi over three years. No institution. No grant funding. No co-founders. No team. A laptop and a clear enough understanding of the problem to build a solution the academic literature does not yet describe.
What is proven: 78.23% directional accuracy, 1.49 Sharpe ratio, 71% bear market accuracy, $0.064 full-stack COGS per user per month, a 107-page research thesis documenting every architectural decision and its empirical basis. These do not require belief. They require reading the documentation.
What is not yet proven: user acquisition, commercial traction, the magnitude of the intelligence scaling gain at scale, and every cross-domain application described here. These require the investment to test.
What I am asking a serious investor to evaluate: given what one person built alone with zero resources, what becomes possible when that person has a team, a registered company, a production deployment, and partners who share the vision?
I believe the answer is significant. I believe this is the beginning of something that matters far beyond cryptocurrency trading. I am looking for the people who see that, and who want to build it.
The complete technical documentation — the full 107-page research thesis, production monitoring data, architecture specifications, and inference economics — is available at the button below. The full 107-page thesis — every architectural decision, every empirical measurement, every inference economics calculation — downloads instantly. I will share additional materials under NDA with any serious investor upon request.
I Want to Talk to the Right People
If you are serious about what this is and what it becomes — not just the trading returns, but the architecture and the mission — reach out directly.