Lucy · Autonomous quant trading
Lucy AI Markets
Lucy AI Markets is a from-scratch trading engine built around a simple discipline: a signal is only trusted when it survives gate-verified data, measured costs, and honest backtests. A council of specialist agents proposes, a single risk gate decides, and one execution door acts — so “approved X, placed Y” is unrepresentable by design.
Systematic traders and allocators who want the process, not a black box.
The problem
Backtests lie by default
Almost any strategy can be made to look profitable. Mixed-interval data invents edges that never existed. Ignore spreads and financing and a losing system prints money on paper. Try a thousand variations and one will shine by luck alone. The industry calls that research; we treat it as the central engineering problem.
How it works
Signal path — proposal to execution
Our approach
The decisions behind Lucy AI Markets
Verify the data before you believe the result
Historical data passes a battery of integrity gates before any strategy touches it. If the bars are inconsistent, misaligned, or silently patched, the run is refused rather than reported. A result computed on unverified data isn't a weak result — it's not a result at all.
Charge the strategy what trading actually costs
Spreads and financing are measured per instrument, not assumed. Fills are modelled honestly, including the unpleasant cases — gaps straight through a stop, entry-bar stop-outs. Most 'edges' die here, which is exactly the point.
Make safety structural, not procedural
Every intent passes the risk gate inside the single execution door. The door takes no verdict argument, so a mismatch between what was approved and what was placed cannot be expressed in the codebase. Guarantees you can't accidentally bypass are worth more than guarantees you have to remember.
Let the broker keep score
Performance is read from the venue, never self-reported by the engine. A system grading its own homework is the oldest failure in this field.
What it does
Capabilities
- Multi-agent council for research, entry, exit, and risk
- Evidence-gated strategies — validated on 30y of cost- and financing-corrected data
- One risk gate, one execution door — safety is structural, not optional
- Broker-truth accounting: performance comes from the venue, never self-report
Under the hood
How it's built
- Python 3.13 · FastAPI · React + TypeScript
- Capital.com execution adapter
- Fernet-encrypted credential store, no plaintext secrets