Manufacturing — 12 months

Engineering Handbook Assistant

Forty years of process documentation, drawing annotations and revision history — scanned, superseded, and scattered across three generations of document systems — collapsed into a retrieval layer that shop-floor staff trust enough to act on, running entirely on hardware with no path to the outside network.

Challenge

The corpus was never built to be searched: scanned drawings with handwritten margin notes, superseded revisions sitting alongside current ones with no reliable metadata to tell them apart, and process documents written in a house shorthand that assumed the reader already knew the equipment. Air-gapped hardware was non-negotiable from day one.

Approach

Ten weeks of ingestion work — OCR tuned for the drawing annotation style, a revision-resolution pass built with engineers who knew the informal versioning conventions, and chunking tuned to preserve the link between a drawing and its annotations. Hybrid dense-plus-keyword retrieval with re-ranking, every answer resolved back to the exact source paragraph.

Outcome

In production for six months as of writing, running entirely on the client's own hardware with no network path out. The evaluation approach rebuilt mid-project from logged real queries is now the default starting point for every new deployment.

Stack

Open-weight LLMHybrid retrievalAir-gapped inferenceCitation grounding