Home Hub

Why This Matters

UK broadband providers are sitting on an unrealised opportunity. They already have hardware in every home, relationships with energy suppliers via the smart meter rollout, and the technical capability to integrate security and home automation. What is missing is the software layer that ties it together and turns it into daily value.

Home Hub explores what that could look like: a single interface built around the gateway that makes connectivity, energy, security, and family controls genuinely useful together rather than separately.

Real-time telemetry: performance, control, and diagnostics

Every device in a Home Hub deployment is a source of continuous telemetry. The gateway publishes its own hardware metrics: CPU load, memory pressure, thermal state, uptime. Each mesh repeater publishes signal quality, channel utilisation, connected client counts, and per-device throughput. The smart meter publishes consumption, generation, export, and tariff state, typically at 10-second intervals. Connected appliances (heat pump, EV charger, solar inverter, battery storage) publish operating state and energy draw in real time. All of it flows over MQTT to a local broker on the gateway, which routes events to the relevant application layer or queues them for cloud synchronisation at sub-second latency.

That telemetry stream is what makes performance management possible. The platform monitors Wi-Fi mesh health continuously: when a node's signal degrades or its client count drops unexpectedly, it detects the pattern before a household notices a problem and either reroutes clients automatically or surfaces a diagnostic to the operator. The same logic applies to connected appliances: deviation from expected energy draw or operating cycle is an early signal of appliance stress or fault, detectable weeks before failure becomes visible.

Control decisions are driven directly from the telemetry stream and execute on the gateway without a cloud round-trip. When smart meter data shows solar export peaking and grid tariff prices are low, it shifts EV charging load in real time. When a demand flexibility signal arrives from the energy supplier, it responds to controllable loads within the agreed latency window. Because the decision logic runs locally, the response time is measured in milliseconds rather than seconds.

Local inference on the gateway is made practical by dedicated NPU silicon from vendors including Qualcomm, MediaTek, Broadcom, and Nvidia. These processors allow quantised models to run continuously at low power alongside the gateway's routing and telemetry functions. Lightweight open-source models (Llama, Mistral, Phi) handle pattern recognition and anomaly detection on-device, exported via TFLite or ONNX Runtime for efficient NPU execution. For tasks that require deeper reasoning or natural language generation, the cloud layer uses larger models (GPT-4, Claude, Gemini) to process anonymised fleet data, retrain inference models, and generate insights that are pushed back to the gateway fleet. The choice of model is always driven by the task: on-device for anything latency-sensitive or privacy-critical, cloud for anything that benefits from scale.

Security is designed into every layer. Input validation and schema enforcement on all MQTT messages prevent malformed or malicious payloads from reaching the inference stack. Model outputs pass through output filtering and confidence gating before triggering any control action. For any cloud-facing model interaction, prompt injection defences and adversarial robustness testing are part of the deployment pipeline, not an afterthought. Household data stays on-device by default. Only anonymised, aggregated telemetry is sent to the cloud training pipeline, and model weight updates are signed and verified before the gateway accepts them.

For storage, high-frequency telemetry goes into a time-series database. That is the right structure for the windowed queries and trend baselines that diagnostics and performance management require. Events, control actions, and state changes go into an event store, preserving the full operational history in sequence. The cloud analytics layer processes aggregated, anonymised data from the fleet to identify patterns and retrain local inference models, pushing updated model weights back to each gateway. The result is a deployment that improves across the installed base over time without requiring individual household data to leave the home.