Project · Financial AI

Trader Koo

A production financial-markets system — multi-agent critic, walk-forward ML, four hard safety guardrails. Paper-trade only, by design.

Financial AI Safety-Critical Multi-Agent Production

The problem

Markets are noisy. ML classifiers overfit. Most personal trading projects amplify risk without serious checks — a candidate model with AUC 0.52 is indistinguishable from luck, but it still tells you to buy.

I wanted a system where every signal passes through gates I would actually trust: regime alignment, volatility threshold, model confidence floor, position sizing capped on dollar risk — not on tier notional. And one where the worst possible outcome is a paper loss.

Architecture

Ingest
yfinance + Finviz + Finnhub WS + Binance WS + FRED + Polymarket. Nightly orchestrated update at 22:00 UTC plus 24-hour cache TTL on heavy endpoints.
Predict
LightGBM walk-forward classifier on 51 engineered features. HMM regime detector (calm/normal/stormy). YOLOv8 chart-pattern detector (foduucom/stockmarket-pattern-detection-yolov8) run batch nightly.
Critic
A Claude-backed critic reviews every candidate setup against regime, VIX threshold, ML confidence, and a 5-strategy signal ensemble. Setups failing any gate are rejected before they reach the paper book.
Sizing
Dollar-risk-budget-based at ~0.5% per trade, anchored to next-day open + slippage, capped by tier ceilings. Wide-stop trades automatically sized down.
Execution
Paper book only. Next-day open entry, slippage, commissions, borrow costs simulated. ADV gate, gap fill, dividend handling, SPY benchmark with dividends reinvested.
Observability
Pipeline health dashboards, per-ticker YOLO event tracking (ok/skipped/timeout/failed), paper-trade equity curve vs SPY, Sortino + Calmar + expectancy gate.
Evals
Walk-forward backtester. Feature-activation checklist before any new gate goes live: sample size ≥ 20, AUC > 0.55 on 5+ folds, calibration curve check, leakage audit.

Safety — the part that matters

This is where the FDE instinct lives. Every guardrail is a hard block, not a soft suggestion. No override-on-high-score escape hatches.

What the numbers say

603+
Tests passing
51
ML features
5
Signal strategies
$0
Real capital at risk

By design. Not for sale, not for advice. The point isn't to make money — it's to demonstrate that LLM-orchestrated decisioning can be built with the same discipline as a regulated trading system.

Tech stack

Python FastAPI APScheduler Claude LightGBM YOLOv8 PyTorch SQLite React 19 Vite 8 TypeScript Plotly Railway