AI Financial Agent — Engineering Case Study
Stealth FinTech Startup · by Yatharth Lakhera
Production-ready AI financial agent built for a stealth FinTech startup — multi-agent RAG over Gmail financial data with >95% classification accuracy, running at just ~$100/month. The 2-month MVP helped the founders close $100k in pre-seed funding.
The challenge
A stealth FinTech founder needed an investor-ready AI financial agent in time for a pre-seed raise — production-grade accuracy, a real conversational UX, and unit economics that wouldn't blow up at scale. Two months on the clock, with a tight infra budget and fragmented financial data scattered across bills, cards, and statements buried in Gmail.
The build
Architected a multi-agent RAG system over Gmail financial data using GPT-4 + Gemini Flash with tool calling, Qdrant for vector retrieval, and Supabase + n8n for orchestration. Agents ingest bills, credit cards, and statements, normalize them into a unified view, proactively flag late-fee risk, and let users 'ask' their finances directly through a natural language interface. Cost-engineered the model and retrieval mix so the entire system runs at ~$100/month at MVP scale.
Impact
- Helped the stealth startup close $100k in pre-seed funding on the back of the MVP
- Production MVP delivered in 2 months — zero to investor-ready
- >95% classification accuracy on heterogeneous Gmail financial data
- Single unified view across bills, credit cards, and recurring payments
- Proactive late-fee detection — alerts fire before charges hit
- Cost engineered to ~$100/month — defensible unit economics for the pitch
Technologies: Python, FastAPI, GPT-4, Gemini Flash, Google ADK, Qdrant, RAG, n8n, Supabase
Yatharth Lakhera — full portfolio