Three stages. The middle one is where the language model lives — and it is metered openly.
Service manuals and bulletins (PDF), parts catalogues, closed work orders with resolutions (export), and — per case — the technician's photo, alarm code, machine model/serial and a one-line description.
Identifies the model and revision, retrieves the relevant manual sections and past cases, reads the photo for visible conditions, and reasons through the diagnostic tree: what to check first, what each outcome means, when to stop and escalate.
An ordered checklist with manual citations, likely parts with catalogue numbers, safety prerequisites up front, a draft work-order summary, and a flag when the case has no manual coverage so a senior engineer is pulled in.
Illustrative output.
Each case reads manual sections, past work orders and one or more photos — typically 50–200k tokens — and a network of 50 technicians runs hundreds of cases a day. Expect 30–200 million tokens a month at scale. Retrieval is indexed once; the model is spent per case on reading and reasoning.
We are onboarding a small number of design partners with real documents and real deadlines. Tell us what you work on and we will set up a workspace with your material.