Spendy turns a bank statement PDF into an actual spending picture: Claude extracts every transaction, categorizes it, and writes a plain-language summary, feeding a dashboard of spend by category, spend over time, and top merchants. Three services do the work — a React dashboard, a Hono API for auth and CRUD, and a Python/FastAPI service that owns PDF extraction and the Claude pipeline — with real attention paid to the security side of handling financial data: in-memory-only PDF passwords, AES-256-GCM encryption at rest, pre-LLM redaction of account numbers, and a 30-day retention window on raw statements.
Read the full write-up in the linked post for the architecture, the privacy decisions, and the loan-tracking Telegram bot that came out of using it on myself.