LLM Engineer — Make Our Claude-Based PDF Extraction Faster & More Accurate

he Glass Market (B2B window & door platform) has a live AI feature: reps snip schedule tables from architectural plan PDFs, and Claude extracts a structured 11-column takeoff. It works — your job is to make it faster and more accurate, with every change measured. Stack: Anthropic Claude API, Supabase Edge Functions (Deno/TypeScript), React. We already capture AI-output vs human-corrected training pairs. Note: this codebase is built and maintained with Claude Code and will keep being edited by it — you must be comfortable working alongside AI-authored code via clean, documented PRs. Deliverables: Eval harness scoring per-field accuracy against our gold data, with a baseline report Measured accuracy improvement (prompt/few-shot tuning + a corrections-feedback loop) Measured latency reduction on both extraction paths Short report: what changed, before/after numbers, what's next Skills: Anthropic Claude API (structured outputs, prompt engineering for document extraction) LLM evaluation experience TypeScript/Deno, Supabase/Postgres Experience with AI coding agents (Claude Code, Cursor)