We are building the first hedge fund where the entire research process — from hypothesis to portfolio — is run by autonomous agents, not analysts.
When a researcher spends 3 months testing a factor and it fails, that failure is lost. The next researcher tests the same thing. The most expensive waste in quant is repeated research.
A failed backtest is treated as a dead end. But why it failed — regime dependency, data bias, cost structure — is worth more than the factor itself. Those insights vanish when the notebook closes.
Each researcher uses their own tools, style, and mental models. When a senior quant leaves, the fund's research capacity drops overnight. There is no shared operating system.
One researcher tests 50–100 hypotheses per year. An AI agent on structured infrastructure can test 10,000. The question is not if — it's who builds the right infrastructure first.
Every experiment — successful or failed — is stored in a structured, queryable knowledge base. Hypothesis, data used, result, diagnosis, and connections to related experiments. When the system encounters a similar hypothesis, it retrieves prior results automatically. This turns research from linear to compounding.
PostgreSQLNeo4jPineconeAn autonomous agent that continuously reads new research, detects market patterns, and cross-references the Research Memory to generate novel factor hypotheses — each with explicit assumptions, testable predictions, and failure criteria.
Claude OpusSSRN / arXivStructured OutputEach hypothesis is automatically translated into a backtesting specification: stock universe, factor logic, parameters, cost model. The experiment runs autonomously. Results stored regardless of outcome. Every experiment is reproducible from its stored spec.
qlibPythonDockerKills bad factors before they touch capital. Out-of-sample testing, regime analysis, turnover stability, transaction cost impact, correlation checks, and decay analysis. Default stance: rejection. False positives are 10× more expensive than false negatives.
OOS TestingRegime AnalysisCost ModelValidated factors combined into a portfolio with systematic risk management. Algorithmic execution minimizes market impact. Humans set risk limits and maintain oversight — not stock picking. AI proposes, human disposes.
Risk ParityTWAP / VWAPReal-time PnLFor the first time, AI can parse a 40-page quant paper and extract the actual testable hypothesis. This unlocks the entire paper → hypothesis → experiment pipeline that was previously bottlenecked by human reading speed.
AI agents can now write, debug, and iterate on backtesting code that actually runs in production. The Experiment Engine depends entirely on this capability — and it only became reliable in late 2024.
The Research Memory requires structured storage of relationships between hypotheses, experiments, and outcomes. Vector DBs + graph DBs + LLM-powered retrieval became production-ready in 2024–2025.
Knows how capital allocation works at the highest level. Understands LP expectations, fund structure, and risk management in practice. Was inside one of the world's most sophisticated investment machines.
Built inside the most advanced systematic trading infrastructure on earth. Knows how real quant research works, where it breaks, and what "good" looks like. Has lived the exact pain point we're solving.
The person who prevents two finance minds from building a system only they can use. Product discipline, system architecture, and the ability to turn a research insight into maintainable infrastructure.