Embedded AI that runs entirely on the device
Real-time event detection within a coin-cell power budget, with raw sensor data that never leaves the device.
- Fits
- Wearables, industrial sensors and connected products
- Platforms
- Cortex-M class microcontroller
- Layers involved
- Cloud and agents, On-device AI, Firmware, Hardware and sensors
- Engagement model
- Fixed-scope project with field support
A representative scenario that shows how we approach this kind of product, rather than an account of a specific engagement.
The scenario
A connected product needs to recognise meaningful events in a continuous stream of sensor data. Sending that data to the cloud for inference adds latency, draws on a small battery and moves sensitive raw signals off the device. This scenario brings the intelligence onto the microcontroller itself, with the aim of detecting events within milliseconds, running for months on a small battery and sharing only derived events.
Objectives
- Detect events on the device in real time
- Run for months on a coin-cell class battery
- Keep raw sensor data private, sharing only derived events
Design targets
- Power
- Coin-cell class budget
- Memory
- Target under 256 KB SRAM
- Timing
- Bounded worst-case latency
- Privacy
- Raw data stays on the device
Our approach
This scenario involves four of the five layers: hardware and sensors, firmware, on-device AI, and cloud and agents, which handles secure updates. The applications layer is kept deliberately thin, because the product's value is delivered on the device itself.
Secure over-the-air model updates
What we build
Signed, versioned model packages with staged rollout, health reporting and automatic rollback, delivered independently of firmware releases.
Key decision
Separating model updates from firmware releases lets detection improve in the field without a full firmware cycle.
A compact model optimised for the microcontroller
What we build
An int8 model distilled from a larger teacher model, trained with quantisation in mind and benchmarked on the target part from the outset.
Key decision
On-device inference meets latency and privacy requirements that are difficult to satisfy with a cloud round trip, and avoids the energy cost of transmitting raw data for every detection.
Inference inside the real-time control loop
What we build
Fixed-size buffers, no dynamic memory allocation on the time-critical path, and sensor sampling synchronised with the inference schedule.
Key decision
Static allocation and a measured worst case make the timing guarantee something that is proven rather than assumed.
Microcontroller and sensor selection
What we build
A Cortex-M part with DSP extensions, an IMU with tight time synchronisation, and a power tree profiled across sleep, sampling and inference states.
Key decision
A capable microcontroller meets the performance target without the cost and power of a separate neural accelerator.
Delivery path
An indicative sequence. The length of each phase depends on the product, and we agree it with you during discovery.
- Phase 1
Discovery and bring-up
Agree the events to detect and the power, memory and timing targets, then profile sensors, timing and power on the target board, so every later decision rests on measured numbers.
- Phase 2
Model development
Data collection, training, distillation and quantisation against the memory and latency targets.
- Phase 3
Firmware integration
Inference in the control loop, worst-case timing validation and power measurement under real workloads.
- Phase 4
Updates and field testing
Signed over-the-air updates, staged rollout and a field pilot with health reporting.
Outcomes the approach is designed to deliver
Raw data never leaves the device
Only derived events are shared, which simplifies the product's privacy obligations and reduces its connectivity needs.
Real-time detection on the device
Events are recognised within a bounded time, without a round trip to the cloud.
Models that improve in the field
Signed updates refine detection over the product's life without touching the firmware.
Key takeaway
Profiling on the target hardware from the outset gives the most reliable timing and power data for every decision that follows.