Multi-agent quantitative research platform that ingests live market data, reasons over it with a team of LLM agents, ranks mispricings with explainable conviction, sizes a risk-disciplined paper book, suggests live options strategies, and measures its own predictive accuracy to improve over time — built end-to-end with only free, real-world data sources.
A from-scratch AI hedge-fund research engine for the Indian market. It fuses real-time data engineering, multi-agent LLM orchestration, quantitative finance (options Greeks, dealer gamma, regime detection), portfolio risk management (VaR, stress tests), and an MLOps-style learning loop (information-coefficient–driven adaptive weights) behind an async FastAPI service and a live terminal dashboard — all running 24/7 on a single machine with zero paid APIs.
Live dashboard — ranked mispricings with conviction, expected return, tail risk, factor attribution and probabilistic scenarios.
| Domain | What I built |
|---|---|
| Data Engineering | Fault-tolerant real-time ETL from 8+ free sources (NSE option chains, FII/DII & F&O participant flows, bulk/block deals, OHLCV, FRED macro, RSS, Reddit). Browser-session emulation + cookie handshake to defeat NSE bot-blocking, exponential backoff, idempotent upserts, freshness heartbeats. |
| Storage | Hybrid OLAP + OLTP: DuckDB columnar warehouse for time-series analytics, SQLite (WAL) for transactional app state, NetworkX for the graph — embedded, zero-ops. |
| LLM / GenAI | Multi-agent orchestration (9 specialist agents) over a provider-agnostic LLM gateway with failover (Gemini ⇄ Groq), token-bucket rate limiting, disk response caching, JSON-mode structured output, and graceful degradation to deterministic rules when quota is exhausted. |
| Knowledge Graph | A live market graph (stocks ↔ sectors ↔ indices ↔ macro) built from curated maps, realised-return correlations, and LLM-extracted relations from news headlines; traversed for second-order contagion signals. |
| Quant Finance | Black–Scholes Greeks, dealer gamma exposure (GEX), PCR, max-pain, IV skew, regime classification, and a 24-strategy options engine (spreads, condors, butterflies, straddles, ratio/synthetic/box) ranked by a direction × IV × gamma decision matrix. |
| Risk Management | Vol-targeted position sizing, per-name & sector concentration caps, drawdown de-grossing, historical VaR / Expected Shortfall, and scenario stress tests (crude, USD/INR, rates, FII-exodus, broad sell-off). |
| ML / MLOps | A self-evaluating learning loop: every signal's factor vector is snapshotted, scored on forward returns, and converted to per-factor Information Coefficient (IC) + hit-rate that adaptively reweights the alpha model — no look-ahead, fully reproducible. |
| Backend / Full-Stack | Async FastAPI REST API, APScheduler market-hours-aware job orchestration, a dependency-free real-time dashboard (vanilla JS + Chart.js), Telegram/email alerting, and one-command 24/7 deployment via a macOS LaunchAgent. |
flowchart TD
subgraph L1["① Data Layer — real-time ETL"]
NSE["NSE API<br/>chains · FII/DII · F&O · deals"]
YF["yfinance<br/>OHLCV · fundamentals"]
MAC["FRED / World Bank<br/>macro"]
NEWS["RSS · Reddit<br/>news · sentiment"]
end
WH[("DuckDB warehouse<br/>+ SQLite state")]
NSE & YF & MAC & NEWS --> WH
subgraph L2["② Knowledge Graph"]
KG["NetworkX · correlations<br/>+ LLM-extracted relations"]
end
WH --> KG
subgraph L3["③ Multi-Agent Team (LLM)"]
A["Flow · Options · Macro · Sentiment<br/>Earnings · Valuation · Regime · Causal"]
end
WH --> A
KG --> A
subgraph L4["④ Alpha + Risk + Learning"]
ALPHA["Alpha engine<br/>expected vs observed"]
SCEN["Scenario engine<br/>bull/base/bear"]
RISK["Risk engine<br/>VaR · stress · sizing"]
EVAL["Eval loop<br/>IC → adaptive weights"]
end
A --> ALPHA --> SCEN
ALPHA --> EVAL --> ALPHA
CIO["⑤ CIO Agent<br/>conviction 0–100"]
SCEN --> CIO
A --> CIO
CIO --> RISK
subgraph OUT["⑥ Delivery"]
PAPER["Paper book<br/>real marks"]
OPT["24-strategy<br/>options engine"]
DASH["FastAPI + dashboard"]
ALERT["Telegram / email"]
end
RISK --> PAPER
WH --> OPT
CIO --> DASH & ALERT
PAPER --> DASH
OPT --> DASH
 Institutional Flow Intelligence — FII/DII cash + F&O participant positioning + bulk/block deals |
 Index Options — PCR, max-pain, dealer-gamma regime, IV skew, OI walls |
 Live Option Trade Ideas — 24 strategies ranked by relevancy, with real strikes, premiums & payoff diagrams |
 Portfolio Risk — VaR / Expected Shortfall, sector concentration & scenario stress tests |
 Learning & Accuracy — forward-return Information Coefficient, hit-rate, adaptive factor weights |
 Market Knowledge Graph — entities & relationships, interactively explorable |
Language · Python 3.12 (type-hinted, dataclasses)
Backend · FastAPI · Uvicorn (ASGI) · APScheduler
Data · DuckDB · SQLite (WAL) · NetworkX · pandas · NumPy · SciPy
GenAI · Google Gemini · Groq (Llama-3.3-70B) · provider-agnostic gateway with failover, rate-limiting, caching
Ingestion · requests (session/cookie emulation) · yfinance · feedparser · PRAW · FRED / World Bank APIs
Frontend · Vanilla JS · Chart.js · vis-network (zero build step)
Ops · launchd 24/7 service · rotating logs · Telegram / SMTP alerting · .env secret management
distributed data pipelines · real-time streaming ETL · multi-agent systems · LLM orchestration · RAG-style retrieval · knowledge graphs · quantitative finance · options pricing & Greeks · time-series analysis · feature engineering · regime detection · Monte-Carlo-free probabilistic scenarios · portfolio optimization · Value-at-Risk · stress testing · backtesting · information coefficient · online learning / adaptive weighting · fault tolerance · rate limiting · caching · graceful degradation · idempotency · observability · REST API design · async I/O
git clone https://github.com/ChethanKMurthy/CTGTerminal.git
cd CTGTerminal
python3.12 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env # add free API keys (all optional — see below)
python run.py # boot warm-up + 24/7 scheduler + dashboardOpen http://localhost:8799 🎉
Every API key is optional — the system degrades gracefully and still runs on pure deterministic logic. Add a free Gemini or Groq key for LLM reasoning, FRED for macro, and a Telegram bot token for phone alerts.
python run.py --once # run one full end-to-end cycle and exit (cron-friendly)
python run.py --web-only # dashboard only, over existing datacp deploy/com.ctg.alpha.plist ~/Library/LaunchAgents/
launchctl load ~/Library/LaunchAgents/com.ctg.alpha.plistctg/
├── data/ # real-time collectors: NSE, prices, macro, news, social, universe
├── storage/ # DuckDB warehouse, SQLite state, NetworkX graph persistence
├── llm/ # provider-agnostic LLM gateway (failover, rate-limit, cache)
├── agents/ # 9 specialist agents (flow, options, macro, sentiment, ...)
├── engine/ # knowledge graph, alpha, scenarios, regime, risk, evals,
│ # quant analytics, 24-strategy options engine
├── portfolio/ # risk-sized paper-trading book with real marks
├── alerts/ # Telegram + email dispatch with dedup
├── scheduler/ # market-hours-aware job orchestration + pipelines
└── web/ # FastAPI app + real-time dashboard
Defeating NSE bot-blocking & building a resilient ingestion layer
NSE's public JSON endpoints return 403/401 to naked requests and block datacenter IPs. The client emulates a browser: warms cookies via the homepage + option-chain page, sends a full header set (and deliberately omits brotli, which silently corrupts JSON), throttles requests, rotates the session on auth failure, and retries with backoff. Endpoints also change shape over time (the option chain moved from /option-chain-indices to a contract-info → option-chain-v3 two-step) — the client adapts. Every collector is idempotent (INSERT OR REPLACE on natural keys) and writes a freshness heartbeat surfaced on the dashboard's System tab.
The self-learning eval loop (the MLOps bit)
Static factor weights overfit and decay. Instead, every cycle snapshots each candidate's full factor vector + entry price. Once genuine forward prices exist (no look-ahead), the loop computes per-factor Information Coefficient (correlation of factor value with forward return) and the composite hit-rate, then blends those into adaptive weights the alpha engine consumes next cycle. The model literally tilts toward whatever is predictive right now and is fully reproducible from the warehouse.
Free-tier-proof LLM gateway
A single gateway abstracts Gemini and Groq behind one interface with provider failover, a sliding-window token-bucket rate limiter, a TTL disk cache keyed on (model, prompt), JSON-mode parsing with repair, and cooldown-based circuit breaking: when a provider hits its daily token cap it's benched for 30 min and the whole system falls back to deterministic rules — no crashes, no 20-second retry storms, no dropped cycles.
Risk-first capital allocation
Signals don't become positions blindly. The risk engine vol-targets each name (size ∝ conviction ÷ realised vol), enforces per-name and per-sector caps, de-grosses on drawdown, scales gross by regime, then reports 1-day historical VaR/ES and marks the book against five macro stress scenarios so no single shock is uncomfortable.
Research & educational project. Not investment advice. All trading is simulated (paper). Free market data is delayed; option premiums are indicative, not executable quotes. Options and futures carry high, sometimes unlimited, risk.
MIT — see LICENSE.
Built with a focus on real data, sound quant, and production-grade engineering.