IMAGE
An autonomous AI paper-trading system — a four-LLM committee debates every market headline in real time, then trades only when conviction, risk, and quality gates all clear. Running unattended for weeks with near-zero errors.
Most trading bots pattern-match headlines and fire orders. Sentient Trader is fundamentally different — it runs a sequential four-agent AI debate where each persona reads and reacts to what the previous agent actually said, producing genuine disagreement rather than averaged noise.
Built on LangGraph and Groq-hosted LLaMA 3.1, the system ingests live market news via Alpaca WebSocket, routes signals through Redis Streams, and processes every headline through a full committee debate. The momentum trader sets a thesis, the value investor challenges it with fundamentals, the risk manager stress-tests both sides, and the portfolio manager synthesises the final call — all with complete decision traces stored for replay.
A pure-Python risk gate then evaluates five independent conditions before any trade executes. The system has been running autonomously since May 2026 — processing thousands of signals, executing risk-gated paper trades via Alpaca, and persisting every decision trace for full auditability. Safety boundary: paper trading only, educational use only, not financial advice.
Real-time signal feed, PnL equity curve, system health, and full agent decision traces — no login required for read access.
OPEN DASHBOARD →Every decision is replayable as an interactive pipeline — news → pre-screen → market context → 4-agent debate → risk gate → execution.
REPLAY GOOG SIGNAL →Submit your own ticker and headline to watch the full committee debate unfold in real time. Simulated signals are safely blocked from order execution.
TRY SIGNAL INJECTOR →Sequential, not parallel. Each agent reads and reacts to the previous agent's actual reasoning — producing a real argument, not independent guesses in a vacuum.
LLM call #1 — analyses price action, momentum indicators, and short-term directional bias. Sets the opening thesis that subsequent agents react to.
LLM call #2 — reads the momentum trader's actual take, then evaluates the headline through a fundamentals lens. Disagreement here is substantive, not coincidental.
LLM call #3 — stress-tests both sides. Identifies blockers, assigns risk levels, and flags concentration or circuit-breaker violations before synthesis.
LLM call #4 — the synthesiser. Reads the full three-agent debate, issues BUY / SELL / HOLD, and writes the decision trace. The final pure-Python risk gate decides execution.
Not parallel guesses — each agent reads the previous agent's actual reasoning. The value investor reacts to the momentum trader's thesis, producing genuine disagreement.
Every signal persists the complete four-agent debate, risk gate verdict, and final call as JSONB. Open any signal on the live dashboard to replay the reasoning.
Pure-Python quality scorer evaluates every headline before spending LLM budget. ~45% of signals resolved with zero LLM calls — low-quality noise filtered for free.
Five independent conditions must all pass in pure Python: sentiment threshold, confidence calibration, article quality, execution plan, and position risk — no LLM in the loop.
ModelRouter handles rate limits, quota exhaustion, and outages automatically — cascading across Groq and OpenRouter providers within the same signal processing.
Every decision is replayable as an interactive pipeline — news → pre-screen → market context → 4-agent debate → risk gate → execution — reconstructed from stored traces.
RSI, SMA, EMA, MACD technical indicators pipeline with volume ratio and momentum scoring feed directly into agent context for data-grounded LLM reasoning.
15+ institutional-grade trading features — circuit breaker, bracket orders, trailing stops, concentration limits — all toggleable via database config, zero redeployments.
At-least-once delivery via Redis Streams + consumer groups. Nothing is lost if Redis goes down. 4-layer deduplication from ingestion through agent cache.