Personal Finance Assistant
A conversational AI for portfolio queries with three-layer intent routing and grounding validation.
A proof-of-concept showing how to build trustworthy conversational AI for finance. Ask natural questions like "How did my portfolio do YTD?" or "What's my AAPL position worth?" and get grounded answers with full observability into every routing decision and tool call.
The Problem
Building conversational AI for finance is tricky. Users ask questions in unpredictable ways, you need to call the right data APIs without wasteful fan-out, and the LLM must cite its sources accurately. One hallucinated number could mean real financial harm.
The Solution
Three-layer intent routing: fast rule-based matching first, then a lightweight PyTorch classifier, with LLM fallback for ambiguous queries. Every response includes grounding validation to ensure the LLM cites actual tool data. Full tracing shows exactly how each query was processed.
Interesting Details
- •Three-layer routing: rules (fast) → PyTorch bag-of-words (lightweight) → OpenAI (fallback)
- •11 intent types covering positions, activity, quotes, performance, transfers, and general finance facts
- •Grounding validation ensures LLM responses cite actual tool sources, with auto-retry on mismatch
- •Full observability with per-request tracing, latency breakdown, and token cost estimation
- •Smart tool minimization calls only necessary APIs (never fan-out for simple queries)