Ambien — Smart Home Automation & Energy AI

A smart home hardware brand sold great devices but had a fragmented app experience — customers needed separate apps for lights, thermostat, and security. Ambien unified everything under one roof, added cross-device automation, and layered in an on-device ML model that learns each household's patterns without sending behavioural data to the cloud — a key differentiator in a privacy-conscious market.

• 1

  40+ device brands use different protocols (Zigbee, Z-Wave, Matter, Wi-Fi REST) — a universal integration layer was non-trivial.

• 2

  Privacy was a core promise: learning occupancy patterns while keeping all behavioural data on-device required efficient on-device ML, not cloud inference.

• 3

  Automations that are wrong (lights off while you're still in the room) are worse than no automations — confidence thresholds had to be conservative.

• 4

  Multi-user households have conflicting preferences — the model needed to learn per-person patterns via presence detection, not household averages.

The integration layer runs a local home hub (Raspberry Pi or a NAS) that speaks all four protocols and exposes a unified MQTT broker. The Flutter app communicates with the hub over the local network (with end-to-end encrypted cloud relay as fallback). On-device ML is implemented using TensorFlow Lite: a lightweight LSTM model trained locally on the device using federated-learning-style local fine-tuning — no raw behavioural data ever leaves the home. Person identification uses Bluetooth Low Energy proximity from paired phones, so the system knows which household member is in which room. Automations are only triggered above an 85% confidence threshold; below that, the app suggests rather than acts.