on-device.ai puts lightweight edge agents on an OEM's installed base — instruments from three, five, or ten years ago — with no firmware rewrite and no hardware change. Service teams see fleet health, catch drift before failure, and resolve more cases remotely.
Talk to an ExpertOn the wire
Models trained on the instrument's own operating data flag calibration drift, thermal variance, and component wear before they become a field failure.
Calibration and service are scheduled from real usage, not a fixed calendar. The alert goes out before the unit leaves spec — the expectation a lab, a weld cell, or a heavy-equipment fleet already has.
Cycle counts, error codes, and environmental conditions for every connected instrument. A service engineer can open any device in seconds instead of waiting for a site visit.
When an event happens, the platform writes a structured record with timestamps, sensor context, and resolution history for the quality file — a lab audit, a weld procedure record, or a fleet work order.
It lives with the fleet. Overnight it reads live telemetry, the service manual, and the last work orders on that serial number. When weld current climbs, an optic loses alignment, or a GC baseline drifts, it drafts the fault, the likely part, and the procedure — then waits for an engineer to accept, edit, or reject it.
It runs the case. It opens the ticket, checks parts and the customer's service window, drafts the note to the site, and hands the field tech a brief that already has the history. When the visit is done, it writes the closeout: what changed, what was verified, and what the quality file needs.
Procedures, fault codes, exploded diagrams, and resolved cases are the memory. The agent cites the step it used. It does not invent a torque value or a calibration limit. If the manual is silent, it says so and escalates instead of guessing.
The agent prepares. A named service engineer releases. Permissions are scoped: read the fleet, draft the diagnosis, never change a method or ship a part without approval. Every action is written to the same audit log as the instrument event.
Every recommendation is graded against what actually happened. Did the engineer accept it? Did the part fix the fault? Did the instrument come back inside spec? Accepted, edited, and rejected cases become the next exam. A new model does not reach a live instrument until it passes the OEM's own resolved cases.
Public benchmarks do not know your error codes. The eval set is built from real serial numbers, real overrides, and real closeouts. Replay checks three things: grounding (did it cite the procedure), tool use (did it look up the part), and outcome (did the fix hold). A regression stops the release.
A service case is a week, not a question. Alert at 2 a.m., remote session Tuesday, part in transit Wednesday, install Friday, verification run Monday. The agent keeps that state across shifts and time zones. When new telemetry arrives mid-case, it re-plans. It does not start over, and it does not forget Friday's check.
Long horizon means the agent is allowed to do nothing until the instrument speaks again. It notices a self-recovery and closes the loop. It notices a verification run still out of limit and reopens the case with the earlier evidence attached. The quality record is written once, from the whole arc.
Oven, inlet, and detector. The analytical line we connect. Calibration follows real use, not a fixed calendar.
Power sources, wire feeders, and torches already in the cell. Arc stability and cooler faults show up before the line stops.
Excavators, cranes, and the machines that still send a technician when a controller or a hydraulic circuit faults.
Microscopes, optical inspection, and the other bench instruments a lab still services one serial number at a time.
A lightweight edge agent goes on the instrument's existing OS or gateway. No firmware rewrite, no hardware change, including units shipped years ago.
Structured telemetry moves over MQTT or REST with TLS, including air-gapped and low-connectivity sites. The device buffers locally so a dropped link does not lose data.
Data lands in the OEM's service system, LIMS, or cloud — or in a managed cloud — through pre-built connectors. Typical time to a first dashboard is two weeks.
Encryption, role-based access, and an immutable audit log. Lab lines can meet 21 CFR Part 11, ISO 13485, EU Annex 11, and ALCOA+. The same log covers a weld procedure record or a fleet work order.
Condition monitoring, calibration alerts, and failure-mode analysis on dashboards the service team and the end customer can both use.
Health scoring, remote-diagnostic SLAs, and predictive calibration become paid tiers on top of the existing service plan — recurring revenue without a new headcount model.
A roadmap session maps your installed base to a service model, the agents it needs, and the loops that keep them accurate.
Schedule a Roadmap Session