DeadZone: the field technician's AI that diagnoses faults with no signal at all
The problem
Field work happens exactly where signal does not: basements, plant rooms, lift shafts, rural sites. Cloud maintenance apps stall there, so crews go back to paper and retype the job later, and a technician staring at a broken pump still has to call the office or drive back for the manual. Existing offline modes only cache forms for later entry — none of them diagnose anything.
The idea
Why now
Google AI Edge Gallery shipped fully offline speech recognition and synthesis on-device on 10 November; MiniMax-H3-Image edits from up to 16 reference images, so a photo of a broken unit can be compared against healthy references; and decision models like Kanana-2-v-decide answer in 0.04 seconds, making on-device triage cheap enough to run on a phone.
What it combines
AI Edge Gallery's offline speech plus H3-Image's reference-based image editing plus a Kanana-class on-device decision model. Speech alone is dictation, image alone is a camera, and a decision model alone is a form — together they are a technician who can hear, see, and decide with no signal.
MVP
Weekend MVP: one equipment type (a common HVAC pump), a phone app that takes a photo, compares it against three reference images, and returns the top three likely faults from a local decision model with the fix steps. Skip: speech input, multi-equipment support, and CMMS sync.
How it makes money
The moat
The offline diagnosis corpus is the advantage: fault photos, symptom phrases and fix procedures per equipment type accumulate per customer and cannot be scraped or bought. Cloud-first incumbents such as ServiceNow and Salesforce would have to rebuild their sync-first architecture around on-device inference, and their business model sells cloud seats, not offline capability.
Distribution
B2B2C: sold as an add-on inside CMMS platforms like MaintainX and FieldPulse, and directly to facilities and field-service operators. The wedge is teams whose sites are basements, mines, farms and offshore rigs — the places cloud apps go dumb.
Why it wins
MaintainX and FieldPulse sell offline modes that cache work orders for later entry; DeadZone diagnoses on the spot with no cloud at all. It is the difference between a clipboard that works offline and a colleague who works offline.
Risks
On-device image models are heavy — H3-Image needs roughly 135GB of downloads including its text encoder — so phone-class deployment needs serious distillation. The MVP de-risks it by testing one distilled model on one device class before promising a full equipment catalogue.
Build it with
- Google AI Edge Gallery: offline speech recognition and synthesis on-deviceFully offline speech recognition and synthesis on the phone — the voice interface with no signal.
- MiniMax-H3-Image: a 33B open-weight image model that renders 16MP stills in 6 stepsReference-image editing: compare a photo of broken equipment against healthy references.
- Kanana-2-v-decide: Kakao's open decision model answers in 0.04 seconds0.04-second on-device decisions: which fault, which fix, which part.
Repo to start from
deadzone: an offline-first field diagnosis app — on-device speech, reference-image fault comparison, and a decision model that names the fault and the part.
Evidence
- Google AI Edge Gallery: offline speech on-device
- MiniMax-H3-Image on Hugging Face
- MaintainX pricing: Essential $20 per user per month
- A field-service app that works offline (dead zones cost data and trips)
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.