Enterprise AI Data Analyst

A Multi-Modal Enterprise Agent that answers business questions by combining structured SQL data with semantic search over financial documents.

Architecture

1. Vector ETL Pipeline — Ingests 10-K/financial documents, cleans text, splits into semantic chunks using SemanticChunker, embeds with all-MiniLM-L6-v2, and stores in Qdrant with metadata (company_name, document_year).
2. LangGraph ReAct Agent — Two tools:
   • execute_sql: queries a SQLite database of financial metrics (revenue, net income, EPS) with error-recovery.
   • search_vector_db: semantic search over Qdrant for qualitative information from reports.
3. FastAPI + Docker — Wraps the agent as a REST API (/ask), containerized, with response time logging for FinOps tracking.

Companies covered

Apple, Microsoft, Google, Amazon, Tesla, Meta, Nvidia, Samsung, Netflix (2023 data)

How to run

1. Start Qdrant (vector database)

docker compose up -d

2. Set up environment

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Create a .env file:

3. Run the ETL pipeline (one-time)

python etl/pipeline.py

4. Run the API

uvicorn api.main:app --reload --port 8080

5. Or run with Docker

docker build -t ai-analyst .
docker run -p 8080:8080 -e QDRANT_HOST=host.docker.internal --add-host=host.docker.internal:host-gateway ai-analyst

6. Test

curl -X POST http://localhost:8080/ask \
  -H "Content-Type: application/json" \
  -d '{"question": "What was Tesla revenue in 2023?"}'

Tech Stack

• LangGraph (ReAct agent)
• Groq (Llama 3.1 8B) — chosen for free, reliable inference after Gemini quota exhaustion
• Qdrant (vector database)
• HuggingFace all-MiniLM-L6-v2 (embeddings)
• SQLite (structured financial data)
• FastAPI + Docker

Team

Mohan Balaji Rajaram and Ezekiel

Vedio

https://1drv.ms/v/c/5cc8b40456e8a390/IQCUEtgjMdfmS61Gy7fM-WUSAdOxw9qHAAGxMAInAmqqNyQ?e=iQHkZq