Board, firmware, on-device AI, cloud and app — built by one team, on one contract, in China. You leave with working units you can put in an investor's hands.
Not a render and not a deck. Hardware someone can pick up, switch on, and form an opinion about.
Enough to demo in two cities at once and still have a spare for the one that dies backstage.
Documented and buildable by your own engineers on day one. Not a binary blob and not a surprise at handover.
Device registry, update pipeline and data path, deployed into your cloud account rather than ours.
iOS and Android builds on TestFlight and the internal track, ready for a beta group next morning.
Costed at 1k, 10k and 100k units, with second sources already identified for the parts that will bite you.
Thermals, battery life, model accuracy in a real room — and the parts of the concept that did not survive contact with hardware.
Most of this job is finding out what does not work.The schedule holds because we fail early and in public — a demo every Friday, on the actual hardware, including the weeks when it barely runs.
The clock starts when the spec is locked, not when you first email us. That is why Sprint 0 comes first.
Chip selection, the on-device versus cloud split, a bill of materials with real quotes, and the certification path.
First assembly back from the line. Power rails and peripherals verified, firmware skeleton talking to all of it.
Model quantised and running on target. Wake word, latency and power budget tuned. Cloud and app wired in.
Printed or machined enclosure, full integration, and the rough edges filed down until it survives being handled.
Five units in a case, a demo that holds up on stage, and an honest account of what breaks at volume.
Most AI hardware stalls in the gaps between vendors. The firmware house blames the ODM, the ODM blames the model team, and the launch slips two quarters.
ESP32 through RK3588. Bring-up, drivers, RTOS and Linux, power budgets.
Quantisation, NPU deployment, wake word, streaming speech, latency work.
Device management, OTA, model orchestration, cost control, data residency.
Schematic, layout, mechanical, enclosure, EVT-ready design files.
Pairing, provisioning, and the app your beta users will actually judge you on.
Eight weeks is not us working faster. It is nobody waiting on a handover. Chip selection and model selection happen in the same meeting — and on a device with a fifteen-dollar bill of materials, that is the meeting that decides whether the product can exist at all.
Every stage credits into the next. You will know the number before you write the first email.
Sprint 0 is paid work, and worth doing even if you then hire someone else. Two weeks of honest costing is cheaper than eight months of discovering the product cannot be built at your price.
A paragraph is enough. You will hear back from an engineer within one business day — either to book a scoping call, or to say plainly that this is not a good fit for us.
Send the brief