Our work · INTELLIGENT DEVICES

Models that live on the wearable

A sports-technology venture (in stealth)

ON-DEVICE MODEL EVENT TRAINING PLATFORM 4G TELEMETRY MODELS BACK · OVER THE AIR DETECTION AT F1 0.9912 · IN A MODEL THAT FITS ON THE SENSOR
  • On-device event detection at F1 0.9912
  • A research fleet streaming telemetry from the field over 4G
  • Over-the-air model and firmware update pipeline, built and operated by us
From sensor to model and back

The constraint: detect and measure fast athletic movements from a wearable, in the field, where connectivity is patchy and compute is tiny.

Our answer, in an R&D program that is still under way: gradient-boosted models trained in Python and compiled to C to run on the device itself - event detection at F1 0.9912 - plus a research fleet streaming sensor data home over 4G, over-the-air firmware updates through the companion apps, and a training platform that keeps every model reproducible. Latest trick: a model deployed into the sensor silicon itself, classifying before the main processor even wakes.

This platform came to us mid-flight, too. We took it over from the venture’s previous developer and migrated it live: production endpoints cut over with zero downtime, every dataset moved non-destructively. Since then the run-discipline has caught up with the research - an automated harness now tests every firmware release across cellular, Bluetooth and over-the-air interfaces, and automated hyperparameter tuning cut the flagship model’s footprint by 64 per cent without giving back accuracy. Even ground truth gets engineered here: high-speed cameras, radar and body-worn sensors, synchronised to within a single frame.

Bring us the hard problem