Building an autonomous research organisation
AAQuant is an AI-native quant fund. Birchblue built its operating system: the agent orchestration, the governance layer, the statistical gate stack, and the cost controls that let it run unattended.
This page covers the engineering. For the fund itself — strategy, data moat, and investor information — see aaquant.ai.
The engineering problem
Quantitative research is human-serial and therefore expensive. A researcher takes a hypothesis, builds it, backtests it, argues for it, and usually discovers it has no edge — after weeks of work. The question was whether that discovery loop could be delegated to a governed agent workforce without lowering the statistical bar.
What we built
- A six-stage governed state machine carrying every hypothesis from ideation through research, development, validation, and paper deployment — unattended.
- Tiered model routing across five classes. Every feature declares the model class its work requires and the runtime enforces it. Mismatches between intended and actual model are flagged automatically.
- A tamper-evident audit chain. Every state transition, gate verdict, and model call written to an HMAC-SHA256 chain with model, tokens, cost, and prompt hash.
- A statistical gate stack — deflated-Sharpe critic at 90% confidence, correlation gate, four-layer decay detection with false-kill guards.
- Self-healing operations. Supervisor daemon with per-step retry, dead-letter, and escalation; liveness probes; daily machine-written postmortems.
Operating results
| Research cost per terminal verdict | $0.88 — against analyst-months at $10k–100k+ |
| Ideation to verdict, median | 3.66 days — against 3–12 months per strategy |
| Unattended tick success, 30 days | 98.4% |
| Failures resolved without a human | 97.8% of 1,645 |
| Model calls, individually ledgered | 13,228 across 29 governed features |
| State transitions, all audited | 5,041 |
These are pipeline-operations figures. AAQuant is in paper trading and has no live track record; Birchblue publishes its operating metrics, not performance claims.
Why it matters to your project
Quant research was chosen as the proving ground because the feedback is unforgiving and the compliance bar is real. The lifecycle itself is domain-independent — anywhere agents can be staffed, verified, and audited like an engineering team, the same governance applies.
That is what Birchblue installs: the Agentic Development Lifecycle, explicit decision rights, tamper-evident audit trails, and hard cost ceilings. AAQuant is the evidence we run it ourselves first.
Project Info
- Category: IT Technology
- Client: Birchblue (in-house)
- Location: USA
- Completed Date: 2026