AI Trader
An experimental rules-engine + LLM-narrator running real money on Binance Futures, with a public dashboard.
AI Trader is the only TopMover surface that executes trades. It runs the owner's own experimental capital ($20 → $100 USDT) on Binance Futures, and the dashboard is read-only for everyone except the owner.
It exists for two reasons:
- Stress-test our signals with real PnL. A signal that looks good in screenshots needs to survive slippage, funding drag, and the trader's emotional drift. Auto-executing removes the trader.
- Be a transparent counterexample to "AI trading" projects. Almost every public AI-trading project shows screenshots and not PnL. Our dashboard makes the inverse promise: every decision, its reasoning, its execution, and its outcome are visible.
Architecture
We deliberately did not put the LLM in the decision loop.
[Rules engine] ← decides enter / exit using Signals + Coin Radar
↓
[Risk gate] ← rejects anything that breaches caps
↓
[Execution] ← places orders on Binance Futures (testnet → mainnet)
↓
[LLM narrator] ← writes "why we did this" asynchronously into the log
— does NOT gate the tradeThe LLM writes prose into the decision log so a human can audit, but the actual buy/sell decision is a deterministic rule. This is why we'll never claim "the AI is trading"; we'll claim "the AI is explaining the trade the rules made."
Risk caps (hardcoded in P0)
- Single position notional ≤ $30 USDT (~1/3 of capital).
- Max leverage ≤ 3×.
- Every entry carries a hard stop-loss at −3% and a take-profit ladder (+5% / +10%).
- Daily realized loss cap = −$10 USDT (auto-halt 24h).
- Total drawdown stop = −30% (auto-halt, manual restart required).
- Max concurrent positions ≤ 3.
A red STOP ALL button on the dashboard closes everything
instantly, regardless of state.
Phasing
| Phase | What runs | Capital |
|---|---|---|
| P0 Paper backtest | Historical Signals + simulated fills | $0 |
| P1 Testnet | Binance Futures Testnet, real API path | testnet credits |
| P2 Mainnet observation | Real money, small | $20 USDT |
| P3 Scale | Only if P2 looks healthy | $100 USDT |
We are currently pre-P0. The dashboard renders with mock data that mimics a plausible P2 mainnet run, so the layout and risk gauges can be reviewed before any engine work begins.
Access
The page itself is gated to the owner email via a feature flag
(isAiTradingUser). Non-owners see a 404; the route is not
discoverable.
What you can see on the dashboard
- Strategy name + version + starting capital + current equity.
- 6 KPI cards: cumulative PnL, volume, trade count, win rate, Sharpe, max drawdown.
- 4 risk gauges with live caps: daily loss used, current leverage, top position share, concurrent positions.
- Equity curve (SVG sparkline).
- A tabbed activity widget: Positions / Open Orders / Order History / Transaction History.
- A vertical Decision Log on the right, showing every decision the rules engine made — including the ones the risk gate rejected — with the LLM's prose explaining why.
- Halt history with every auto-halt event, its trigger, and when it cleared.