Ultrafire AI

We built Ultrafire AI for Ultrafire as a private, internal knowledge tool — one place where their team can ask a question across the company's complete library of service manuals and fire safety standards, and get a direct answer with the clause and page it came from.

It's closed by design. There's no public sign-up: every account is provisioned by an Ultrafire administrator, and the entire document library stays inside their own instance.

The problem. Ultrafire's institutional knowledge lived in hundreds of pages of service manuals and NZ standards. Finding the right clause meant knowing which document to open and where to look — knowledge concentrated in a handful of experienced heads. We set out to make that library answerable in plain English, without ever asking anyone to trust an answer they couldn't verify.

Retrieval built for technical documents. Generic document-chat tools embed everything and run a vector search. That fails on this material, because technical content is full of exact tokens — NZS 4541, clause 6.3.2, a specific part number. Semantic search is strong on concepts and weak on identifiers; keyword search is the reverse. Ultrafire AI runs both and fuses the results, so a precise clause lookup and a vague conceptual question both land on the right passage.

Answers you can audit. At ingestion we detect clause, table, figure and appendix headings and carry them through as metadata rather than flattening the document into undifferentiated text. That's what lets every answer cite its standard, clause and page — and show the original source text behind it. In compliance work, a confidently wrong clause number is worse than no answer, so the system is built to have its work checked.

It handles real documents. Service manuals in particular arrive as scans of varying quality. The pipeline inspects each upload and decides for itself: clean digital files are read directly, poor scans are OCR'd and deskewed automatically, and staff can force OCR on stubborn ones. Indexing runs in the background with live status, survives restarts, and swaps content atomically so a document is never left half-updated.

Available wherever the work happens. One backend serves an installable web app and a native iOS and Android app, both streaming answers as they're generated — so the same knowledge is available at a desk or on site. Admins get real operational control: upload and organise documents into categories, monitor and re-run indexing, and manage team access.

Built to be run, not just demoed. Passwords are properly hashed, sessions signed, admin functions gated by role, and sign-in endpoints rate-limited. The system refuses to start with weak configuration and errors loudly rather than quietly discarding a rebuilt index — the difference between a prototype and something a business can depend on.