Energy — 6 months
Field Diagnostics Agent
A quantised model running on ruggedised edge nodes in the field, pairing live equipment telemetry with maintenance history to answer diagnostic questions in locations where there is no network to fall back on — and no second chance if the hardware throttles mid-shift.
Challenge
The binding constraint wasn't model size but sustained-load thermal behaviour on ruggedised hardware: a model that ran fine in a cold-start benchmark but throttled the device twenty minutes into continuous field use was not a working deployment.
Approach
A task-specific evaluation set built from representative field queries first, then a mixed-precision quantisation scheme — attention layers held at higher precision, feed-forward blocks quantised more aggressively — calibrated against domain-specific telemetry and maintenance text rather than generic corpora.
Outcome
Deployed across ruggedised edge nodes with sustained-load performance validated on the target hardware profile, not just cold-start latency — the evaluation approach this project established now shapes how every edge deployment gets sized before hardware is chosen.