Creamy-SmartMoneyAgent

Diverging

Inspiration

Bloomberg Terminal costs $24,000/year. We're students. Institutional-grade signal tools — unusual options screeners, dark pool feeds, alternative data platforms — are either paywalled or shallow. We built our own.

What it does

Enter a ticker — an AI agent correlates six independent signal layers in real time:

Options flow — unusual sweeps and directional positioning
Social buzz — Reddit, StockTwits, X/Twitter, Finviz, Yahoo Finance
SEC filings — 13F institutional positions and Form 4 insider transactions
Price structure — Smart Money Concepts (Order Blocks, Fair Value Gaps, AMD phases)
Institutional crowding — short interest, put/call ratio, holder changes
Dark pool activity — volume absorption patterns inferred from OHLCV

The output is a plain-English verdict with conviction level, targeting the window before retail notices institutional accumulation.

How we built it

FastAPI + SSE backend, React frontend, and a genuine LLM agentic loop (Claude Sonnet / GPT-4o / Gemini Flash) that forms a running hypothesis after every tool result — not a fixed pipeline. A cheaper sub-agent (Claude Haiku / GPT-4o-mini) pre-processes raw social data before the orchestrator sees it.

All data sources are free: yfinance, SEC EDGAR public APIs, Reddit JSON, StockTwits, X via twscrape, Finviz scraping, and Twelve Data OHLCV.

Challenges we ran into

Every "obvious" data source was gated or broken — FINRA TRF files sit behind Cloudflare, CFTC COT URLs return 404, and X/Twitter session cookies expire within hours of login. We replaced each with behavioral proxies or free alternatives, and built volume absorption pattern detection as a stand-in for raw dark pool data.

Accomplishments that we're proud of

A working agentic reasoning loop across six heterogeneous free data sources, with SMC primitives feeding directly into cross-signal LLM reasoning — no expensive APIs required.

What we learned

Any single signal is noise. Cross-layer agreement is the only real signal. And prompt design — requiring a hypothesis after every tool result — is the difference between a script and an analyst.