Sentinel Capital is an autonomous 9-agent investment committee for intraday paper trading on NSE/BSE. Five specialist agents analyse each signal independently (Technical, News & Sentiment, Macro & Policy, Forecasting, Contrarian), plus a Quant Screener that cuts a 233-symbol NSE universe to 50–60 daily candidates using 11 standardised factors and Ledoit-Wolf shrinkage-based diversification. A risk/execution layer enforces GARCH(1,1) volatility guards, a 3% hard stop-loss, and a full Indian-market transaction cost model. Every trade is fully traceable to agent votes and trust scores.
agents.md
Technical Analyst — pure OHLCV math (RSI, EMA crossover, MACD, momentum); no LLM; requires 50+ bars.
News & Sentiment Agent — Gemini filters MoneyControl + Economic Times RSS; skips LLM call if no relevant headlines.
Forecasting Agent — LightGBM 3-class classifier (bearish/neutral/bullish) trained on lagged returns, RSI, MACD, EMA diff, rolling volatility, volume change; 9-bar (45-min) lookahead.
Contrarian Agent — Anthropic Claude reviews all 4 specialist outputs; damps confidence of contradicted agents by 50% × contrarian confidence; casts its own vote as 5th input.
Quant Screener — applies 11 standardised factors (momentum, Kaufman efficiency ratio, ATR/ADX breakout signals) with Ledoit-Wolf shrinkage-based diversification to cut 233-symbol NSE universe to ~50–60 daily candidates; sector cap 30%, correlation cluster cap 3.
consensus.log
The Dynamic Trust Framework avoids majority voting. Agent influence = confidence × trust_score × context_relevance, normalised across all agents. The Directional Confidence Score (DCS) is the signed weighted sum: +confidence for BUY, −confidence for SELL, 0 for WAIT. If |DCS| < 0.15 the system waits; otherwise allocation = min(|DCS| × leverage, leverage). Trust scores are Laplace-smoothed hit rates updated one cycle in arrears.
engineering.log
Delta-sized execution — only trades the difference against the current position; prevents compounding exposure on persistent signals.
Complete audit trail in SQLite — every cycle evaluation (including no-trades) records agent votes, debate output, influence breakdown, consensus, and risk verdict.
All LLM-backed agents degrade gracefully — missing API key or rate limit removes that agent's vote for that cycle only; never crashes the pipeline.
Replay mode uses the exact live pipeline — cached OHLCV bars fed through the same orchestration code, so replay is directly comparable to live.
status.log
Active development as of July 2026. Opportunity Discovery wired; continuous live session and manual mode with suggestion cards both implemented. Known gap: Discovery output does not yet feed the watchlist automatically (requires manual POST /discovery/run). No live brokerage execution — paper trading only; Breeze API session token requires daily manual refresh per SEBI browser-login requirements.