Case file · Trading
The Event-Trader Fleet is an autonomous crypto trading system. An LLM proposes trading strategies; each one runs a gauntlet — backtest, then paper, then live — before it is allowed near real capital. The fleet has generated thousands of candidate strategies. Almost all of them die in the first gate, which is the entire point. The interesting work is not generating ideas; ideas are cheap. It is building the gates strict enough that only the survivors get funded.
The pipeline runs unattended. Strategy generation, evaluation, promotion and retirement happen on a loop, with the human reviewing the survivors rather than babysitting the candidates.
Why it exists
Markets are an honest grader. A product roadmap can be wrong for a year before anyone notices; a trading strategy is wrong by Thursday and the loss is in the ledger. Building a system that has to survive that feedback is the fastest way to learn the difference between a plausible idea and a real one — a discipline that transfers directly back to product.
Shape of the system
┌──────────────┐
│ llm propose │ generates candidate strategies
└──────┬───────┘
▼
┌──────────────┐
│ backtest │ gate 1 — most candidates die here
└──────┬───────┘
▼ survivors only
┌──────────────┐
│ paper trade │ gate 2 — live data, no capital
└──────┬───────┘
▼ survivors only
┌──────────────┐
│ live (small)│ gate 3 — real capital, hard limits
└──────┬───────┘
▼
┌──────────────┐
│ retire loop │ underperformers culled, fleet reshuffled
└──────────────┘
Three hard lessons
1. A strategy that looks brilliant for 36 hours is a sample size of one. Early conviction is the most expensive feeling in the system. The gates exist precisely to outlast it.
2. The gate that says no is usually right. The hardest discipline is trusting a rejection you disagree with. Every override of a gate has, on average, cost more than it saved.
3. Thousands of strategies teach you about overfitting, not about markets. Generate enough candidates and some will fit the past perfectly by chance. The whole architecture is built to distinguish a real edge from a lucky curve.
Stack