LLM extraction of technical data for industrial quoting

Problem

Automate industrial quoting by extracting technical data (dimensions, loads, tolerances, materials) from highly heterogeneous documents, to feed the in-house “E8” calculator without manual re-entry.

Constraint

Documents in varied formats, data that must stay within the European Union, materials expressed as free text to be normalised, and a requirement for field-by-field measurable accuracy.

Approach

• Production extraction pipeline with a vision LLM (Gemini, chosen for EU data residency and its ability to read technical drawings).
• Structured, typed outputs (Pydantic), a versioned data schema, and iterative prompt engineering.
• Hardened extraction with fuzzy matching (rapidfuzz) of a free-text material against a catalogue of 288 client grades, plus physical unit conversion (metric ↔ imperial).
• Evaluation system objectively measuring field-by-field accuracy (in-house comparison engine, then Langfuse).

Result

• Extraction accuracy measured at 85 to 96 % depending on the documents.
• Full-stack application delivered: FastAPI / SQLAlchemy async / PostgreSQL, React 19 / TanStack / Tailwind, containerised (Docker, Dokploy), Clerk auth, S3 storage (SHA-256 deduplication), Logfire observability.