FinMRI — an MRI-style health check for any US stock
A virtual hedge-fund team of AI agents that turns market, news, sentiment and fundamental data into a full-picture risk read — built for understanding what you hold, not chasing signals.
The problem
Most quant tools emit buy/sell signals. That's not what most investors actually need. If you're weighing an equity — or already holding one you don't fully understand — the real question is: where does this stock stand right now, and what are the honest arguments on both sides?
FinMRI is built for that question. The name is the design brief: an MRI for a stock — a diagnostic scan, not a trading tip.
What it does
FinMRI assembles a virtual hedge-fund team of collaborating AI agents and runs a full analysis pipeline on any US stock:
- Analysts gather market, news, social sentiment and fundamental data.
- Bull vs. bear debate — several rounds of structured argument between opposing analyst theses, then a consolidation pass that weighs both sides.
- Trader hand-off — the consolidated view is translated into execution planning.
- Risk team — aggressive, conservative and neutral risk perspectives each review the plan.
- Portfolio manager — a final decision layer produces the comprehensive risk read.
Every stage is visible to the user as the pipeline runs. The output is a full-picture, attributable assessment for medium- and long-term investors — not a black-box score.
How it's built
The agent team is orchestrated with LangGraph, with LangChain and RAG feeding grounded context into each agent's reasoning, behind a FastAPI service layer. It runs as a multi-user web platform — accounts, saved reports, featured analyses — that anyone can open in a browser, not a command-line tool you have to set up yourself.
Design decisions
- Primitives over wrappers. The orchestration is built from the underlying tools rather than wrapping a black box — when something misbehaves in front of a user, you need to be able to see why.
- Debate as a first-class structure. Forcing explicit bull and bear theses before consolidation produces materially more honest output than asking one model for "an analysis."
- A product, not a demo. Multi-user auth, report persistence, rate-aware data ingestion, and clear disclaimers — the unglamorous parts that make an AI system usable by strangers.