Days to Seconds: A Technical Support Brain for the People Who Already Know the Product

A Nordic aluminium building-systems manufacturer with deep technical expertise and a long product catalogue (window systems, facade profiles, fire-rated glazed walls) was carrying a multi-day response time on incoming customer technical questions. The expertise existed inside the team, but it was distributed across people, files, and email threads. Manuals were bilingual. CAD drawings carried the answer to many questions but lived behind a separate system. Each new query was effectively re-research.

• Indexed the company's technical documentation, product slides, and Excel specs into a searchable corpus with Finnish-English-Swedish synonym expansion, on top of Next.js and Supabase pgvector.

• Connected Autodesk APS so the system can read DWG and CAD files, not just text. A lot of the answers customers ask for live inside the drawings.

• Added a verification layer: every AI-generated answer carries page-cited sources, and the technical experts rate answers so the system learns from their corrections.

• Designed two surfaces: a search UI for the technical experts (deep retrieval) and a customer-support panel that drafts email replies with the citations attached.

• Deployed inside the client's own environment with their AI provider of choice, not ours. The vendor stays portable.

Customer technical response time moved from days to seconds. Search time per question went from fifteen minutes to about thirty. Experts spend their time on the hard questions, not on retrieval. The system is trilingual and works inside the client's existing document management. The POC validated the approach; the MVP build is in flight inside the same partnership.