README.md

metadata

title: FinAgent - Autonomous Financial AI
emoji: 📈
colorFrom: blue
colorTo: green
sdk: docker
app_port: 7860
pinned: true
license: mit

📈 FinAgent: Autonomous Financial AI

An asynchronous, multi-agent LLM pipeline that automates quantitative financial research, fundamental document synthesis, earnings-call analysis, and real-time news sentiment scoring — built entirely with open-source models.

🏗️ Architecture

This system uses a deterministic state-machine architecture powered by LangGraph:

1. Planner Agent — Parses the user query and generates a strict JSON task queue.
2. Supervisor — A Python-controlled router that dispatches tasks to specialist agents.
3. Specialist Agents:
   • 🔢 Quant Agent — Live pricing, volume, and volatility metrics via yfinance.
   • 📊 Fundamental Agent — SEC XBRL accounting data + RAG on 10-K filings.
   • 📰 Sentiment Agent — Real-time news headline analysis and scoring.
   • 🎙️ Earnings Agent — Sentiment divergence (Prepared Remarks vs Q&A) and keyword trend tracking from earnings-call transcripts.
4. Summarizer — Compiles all agent outputs into a unified Investment Memo.

🚀 Try It

Type a query in the chat box — here are some examples:

Query	What It Does
"How is Apple's stock doing?"	Quant analysis (price, volume, RSI)
"What are the manufacturing risks in Tesla's latest 10-K?"	RAG retrieval on SEC filings
"What is the market sentiment on Microsoft?"	Real-time news sentiment scoring
"Analyze the latest earnings call for AAPL — compare management tone in prepared remarks vs Q&A"	Earnings-call divergence analysis
"Compare the current stock performance of Microsoft and Google"	Multi-ticker parallel analysis

📚 Pre-Loaded Data

This demo comes with pre-ingested data for immediate use:

• SEC 10-K Filings: AAPL, MSFT, TSLA, GOOGL, NVDA
• Earnings Call Transcripts: AAPL, MSFT (Q4-2024, Q1-2025)

Quantitative data (prices, volume) and sentiment (news) are fetched live — no pre-loading needed.

🛠️ Tech Stack

Component	Technology
Orchestration	LangGraph / LangChain
LLM Inference	Groq API (Llama-3.1-8B-Instruct)
Frontend	Streamlit
Backend API	FastAPI + Uvicorn
Vector DB	ChromaDB
Embeddings	HuggingFace all-MiniLM-L6-v2
Market Data	yfinance, SEC EDGAR API

⚡ Performance Optimizations

This system was deliberately engineered for low-latency response times:

• Parallel Agent Dispatch — The Supervisor routes independent tasks to multiple specialist agents simultaneously (e.g., Quant + Sentiment + Fundamental in one batch) rather than sequentially, cutting multi-agent latency by up to 3×.
• Server-Sent Event (SSE) Streaming — Results stream live to the UI as each agent completes, so users see intermediate progress immediately instead of waiting for the full pipeline.
• Groq Cloud Inference — LLM calls use the Groq API (~200 tok/s on Llama-3.1-8B), eliminating local GPU bottlenecks and delivering sub-second per-agent response times.
• Singleton Embedding Cache — The HuggingFace embedding model is loaded once via @lru_cache and shared across all RAG queries (10-K, earnings, etc.), avoiding repeated 500MB+ model re-initialization.
• Token Budget Tuning — max_tokens is capped at 800 per LLM call to prevent Groq from reserving excessive context window, reducing queue wait times by ~40%.
• Pre-Seeded Vector DB — ChromaDB collections are embedded at Docker build time, so the app starts with zero cold-start ingestion delay.
• Per-Step Latency Tracking — Every agent step reports wall-clock latency in the UI, making performance bottlenecks immediately visible.

📂 Source Code

GitHub Repository