Millwright
Concept
Millwright turns a single phone video of a machine procedure into a structured, illustrated, and searchable Standard Operating Procedure — then lets every technician on the floor ask it questions in plain language and get answers cited back to the exact video timestamp or manual page they came from.
A senior technician films a changeover, repair, or setup once while narrating it out loud. Millwright's pipeline transcribes the narration, segments the video into discrete steps using multimodal video understanding, and auto-generates a draft SOP with annotated frame captures, step text, and safety callouts. An engineer reviews and approves it in one pass before it's published. Existing manuals, tickets, and maintenance logs are ingested into the same corpus, so the knowledge base grows from both new video capture and legacy documents.
On the floor, any technician can then ask, "how do I clear a jam on the Line 3 press?" and get a short answer plus a link to the exact SOP step or manual page it was pulled from — no more hunting through binders or waiting for the one guy who knows.
MVP
- Mobile-first capture app: record procedure video + voice narration, tag machine/line
- AI pipeline: transcription + video segmentation into steps + auto-draft SOP with annotated screenshots
- Simple approval workflow: engineer edits/approves before an SOP goes live
- Bulk ingestion of existing PDFs/manuals (OCR + chunking) into the same searchable corpus
- Floor-facing chat/search interface with cited answers (video timestamp or manual page)
- Single-tenant/private deployment per plant to protect proprietary process IP
Why now
Native visual understanding and million-token context (as seen in this week's Kimi K3 and Gemini Robotics ER2 releases) make it feasible, for the first time, to reliably turn unstructured shop-floor video into structured, citable documentation — and to hold an entire plant's manuals, tickets, and video corpus in one queryable context. Falling agent-inference costs make always-available floor Q&A economically viable per line, not just per enterprise contract.
Unique selling points
- Attacks knowledge loss from a retiring workforce directly, at the point of capture (video), not after the fact
- Corpus compounds with every procedure filmed — the moat deepens with usage, not just seats sold
- Human-in-the-loop publishing keeps liability and safety review intact, unlike a raw chatbot
- Clean, narrow MVP a small team can ship: capture app + video/doc pipeline + cited Q&A, no OT/sensor integration required to start
- Land-and-expand motion: start with one problem line or one retiring expert's knowledge, expand plant-wide
