🚀 Overview
The first tagged build of BitNerve — an AI trading engine that pulls live CoinSwitch perpetual futures data, hands it to an LLM every few minutes, and lets the model open and close positions on a simulated $10,000 account while a dashboard records every decision it makes.
Built over a single weekend, inspired by Alpha Arena.
Live dashboard → https://bitnerve.arjunkrishna.dev
🧠 What's in this build
Trading loop
- 5-minute decision cycle across 6 perps: BTCUSDT, ETHUSDT, SOLUSDT, BNBUSDT, XRPUSDT, DOGEUSDT
- Open-position PnL re-marked against live prices every minute
- The model acts through
createPosition/closePositiontools — long, short, or stay flat - Every invocation is persisted whole: prompt, market snapshot, reasoning, tool calls, finish reason
Market data — CoinSwitch
- 5-minute and 4-hour candles (last 50), funding rate, open interest
- Indicators computed locally: SMA, EMA, MACD, RSI, ATR, volume vs. average
Model
- Any model reachable through OpenRouter, via Vercel AI SDK tool calling
- This build was run on Gemini 2.0 Flash
Dashboard — Nuxt 3, Tailwind, ApexCharts
- Account value curve, live price bar, open positions, completed trades
- The model's full reasoning log, cycle by cycle
Persistence — Postgres via Drizzle
accounts,positions,orders,agent_invocations,account_snapshots- Docker Compose for local Postgres, plus a sync script for Neon
⚠️ What this build is not
Worth being exact, since this tag is a historical marker:
- Paper trading only. CoinSwitch supplies market data; no real orders are ever placed. Positions live in Postgres.
- One account, one model. The dashboard was built for a multi-model leaderboard, but the engine here runs a single account driven by a single configured model.
- Leverage is fixed at 1x. Perpetual futures data, spot-like sizing.
- No risk governor. No enforced stop-loss, drawdown cap, or kill switch — the model is trusted with its own sizing.
Experimental software. Expect rough edges, missing safeguards, and spontaneous AI overconfidence.