🏦🤖 AlphaCrunch: Automated Investment Analyst Agent

AlphaCrunch is a LangGraph-powered conversational AI agent using a QLoRA fine-tuned Mistral-7B-Instruct-v0.2 Finance LLM, trained on virattt/financial-qa-10K SEC Q&A data. It powers RAG (ChromaDB) analysis of S&P 500 10-K filings from the jlohding/sp500-edgar-10k with multi-turn memory and Gradio UI.

📁 Repository Structure

├── .env.sample              # Env vars (Modal token, paths)
├── .gitignore
├── LICENSE
├── pyproject.toml           # UV deps
├── uv.lock
├── README.md                # This file
├── requirements.txt         # Legacy pip
├── data/                    # Chroma_db, fiqa, etc
├── notebooks/               # EDA, experiments
├── outputs/                 # LoRA adapter, checkpoints
│   └── finance-llm-adapter/
├── scripts/                 # ingest_data.py, test_agent.py
├── src/alpha_crunch/        # Core package (see sub-structure)
├── .venv/
└── wandb/                   # W&B runs (LoRA eval)

Core Features

Finance LLM (QLoRA Fine-Tuning)

Fine-tuned on 70% RAG-simulated (context + question → answer) and 30% pure knowledge format; 80/10/10 train/val/test split.

Performed local QLoRa fine-tuning on the Mistral 7B Intruct v0.2 model and deployed to modal, accessed through API (FastAPI).

Key Params	Value	Reason
LoRA rank	r=16, alpha=32	Capacity/VRAM balance (fits RTX 5070 Ti)
Quant	4-bit NF4 + double quant	16GB VRAM
Targets	q,k,v,o,gate,up,down	Near full fine-tune quality
Template	Mistral [INST] chat	Instruction-tuned base

Adapter (~30MB) saved in WandB; eval with BERTScore, LLM-as-Judge, also logged to W&B.

To access the dashboard of the fine-tuning and eval:

• WandB LoRA fine-tunning dashboard
• WandB Evals dashboard

RAG

• Data pipeline: Top 50 S&P by weight (Mar 2026 Slickcharts), CIK-filtered from jlohding/sp500-edgar-10k, 47 unique names post-deduplication.
• Deterministic NER with regex on immutable COMPANY_REGISTRY tuple (500+ S&P names), longest-first matching, and alias resolution (e.g., "Google" → "ALPHABET").
• Hybrid RAG: ChromaDB filters by exact company metadata before semantic search with all-mpnet-base-v2 embeddings; @lru_cache singleton optimizations.

Agent

• LangGraph flow: intent_node → conditional (rag/analyst/help) → Mistral Finance LLM; MemorySaver for sessions.
• Multi-turn memory via LangGraph add_messages reducer and MemorySaver checkpointer with thread_id sessions.
• Intent routing: "rag", "analyst", "help" categories; state hygiene clears retrieved_context to prevent leakage.

UI

• Gradio 6.0 UI: Streaming, custom Anta/Courier fonts, mesh gradient CSS, glassmorphism.

Architecture

Gradio UI → LangGraph (AgentState: messages, intent, context, answer)
           ↓ intent_node (Finance LLM)
      ┌──────┼──────┐
  RAG    Analyst   Help
(Chroma) (Reason) (Info)
         ↓ Answer

src/alpha_crunch/
  agent/
    config.py     # Immutable registry, paths, aliases
    vector_store.py # ChromaDB singleton
    rag_node.py   # LangGraph RAG execution
  state.py       # Pydantic AgentState

User input flows: Gradio ChatInterface → LangGraph graph (intent_node → conditional → llm_node) → Mistral 7B on Modal. gradio

Quickstart

1. Clone and install with uv:

   git clone git@github.com:coffeedrunkpanda/alpha-crunch.git
   cd alphacrunch
   uv sync

Copy .env.sample → .env (add Modal API key).

2. Ingest top 50 S&P 500 filings (CIK-filtered, name-standardized):

   uv run scripts/rag/ingest_data.py

3. Launch Gradio UI:

    uv run python src/alpha_crunch/app.py

   Features custom CSS (loomy mesh gradient, Anta/Courier Prime fonts, glassmorphism). gradio

Example Usage

Query: "What are Apple's main supply chain risks?"
Answer: Apple's supply chain risks include disruptions in manufacturing or logistics, sole-sourcing reliance on certain vendors for critical components, and foreign currency exchange rate fluctuations. These risks could materially affect the Company’s financial condition and operating results.

Sample questions:

• what is asset allocation?

• what is the sec fillings 10k? why does it matter?

• what are the most important concepts in investment?

• What are Apple's main supply chain risks?

• Describe tesla's business.

Limitations & Next

• Single company/query only—no multi-compare yet.
• Add more metadata filtering to improve the accuracy of the agent.
• No information on stock prices. Next: yfinance integration.

Tech Stack

• LangGraph for stateful agents with custom nodes and routing.

• ChromaDB local vector store; uv package management; python-dotenv.

• QLoRA-tuned Mistral-7B-Instruct (finance-specialized on SEC Q&A); Gradio ChatInterface.

• Project Manegement with Linear

Interesting Links

LoRA vs. QLoRA - Pranav Patel

LoRA vs. QLoRA - Redhat

Instruction Tuning