ParseHQ is a solo-built AI invoice extraction tool that pulls line items out of supplier PDF invoices into structured, reviewable data. I built a provider-agnostic extraction layer with Gemini and OpenAI adapters behind one validated contract, so swapping the vision model is a config change, not a rewrite. I treated LLM output as untrusted input, Zod validation at the boundary drops malformed line items and flags the invoice for human review instead of writing unchecked data to numeric columns. The background pipeline runs on a ten-table Drizzle schema, signed-URL retrieval from Cloudflare R2, extraction, then validation passes for total mismatches, duplicates, and column misalignment with exponential backoff on retry. I also closed a quota bypass by moving trial enforcement server-side into a locking transaction that counts committed uploads, not completed extractions.