AI Financial Intelligence Agent

What it is

A multi-tool reasoning agent that chains 5 specialized tools to produce structured investment signals with cited sources — processing 500+ articles per day in under 3 seconds per query.

How it Works

Multi-Tool Reasoning Chain

Uses Claude function calling to orchestrate 5 tools in sequence:

1. News Tool — fetches recent articles by ticker/topic

2. SEC Filings Tool — retrieves 10-K, 10-Q, 8-K excerpts

3. Price Tool — pulls historical and real-time price data

4. Technicals Tool — computes moving averages, RSI, volume signals

5. Sentiment Tool — few-shot classifier for market sentiment

Claude decides which tools to call and in what order based on the query — it's not a fixed pipeline. The final output is a structured investment signal with source citations, generated in < 3s latency.

News Clustering Pipeline

Built an embedding-based clustering pipeline with a 0.92 cosine similarity threshold. Clusters track growth velocity — a cluster that doubles in size in 2 hours signals an emerging story. This reduced false-positive alerts by 45% compared to keyword-based approaches.

Sentiment Classifier

Few-shot classifier (87% accuracy) trained on labeled financial headlines. Backtesting against historical price movements achieved 79% precision on high-impact events — events where sentiment shifted sharply and price followed within 24 hours.

Key Technical Decisions

Why function calling over a fixed pipeline? Financial queries are diverse — "what happened to NVDA last week" vs "compare MSFT and GOOGL on cloud growth" require completely different tool sequences. Letting Claude decide the sequence via function calling handles this naturally.

Why cosine threshold at 0.92? Lower thresholds merged too many unrelated stories. Higher thresholds fragmented the same story into noise. 0.92 was calibrated against 2 weeks of labeled data.

Latency target of < 3s: Achieved by parallelizing independent tool calls (news + price + technicals can run simultaneously; SEC filings only called when query mentions fundamentals). Sequential bottleneck is Claude's reasoning step.

Tech Stack

Python, FastAPI, Anthropic Claude (function calling), OpenAI Embeddings, ChromaDB, PostgreSQL, Redis (caching), yfinance, SEC EDGAR API