Assisted Living

Why it is strategically relevant

Most assisted living technology gets built as a standalone product. This model starts from the opposite direction: the operator already has the connectivity, the gateway, and the relationship with the household. The question is whether that infrastructure can carry a care layer.

Inference at the edge

The gateway runs a stack of lightweight inference models on dedicated NPU silicon from Qualcomm, MediaTek, Broadcom, or Nvidia. Not threshold monitors, not rule engines. These are models trained on population-scale behavioural data in the cloud and deployed as quantised binaries (TFLite, ONNX Runtime) to run continuously on-device, assessing the current state of the home against learned patterns of normal activity. Open-source architectures (Llama, Mistral, Phi) are the foundation: small enough to run at low power on constrained hardware, capable enough for real-time multi-variable inference. Because inference runs locally, it responds in real time and keeps working even when the upstream connection is down.

Sensors and devices publish state changes over MQTT to a local broker on the gateway, where the inference stack processes the incoming event stream continuously. The models draw on motion sequences, door contact timing, temperature gradients, and device state simultaneously to assess what is happening and where. For natural language tasks such as summarising daily activity for a carer or generating plain-language alerts, the cloud layer uses larger models (GPT-4, Claude, Gemini) against anonymised context. The on-device models handle everything latency-sensitive or privacy-critical. The cloud handles tasks that benefit from scale and reasoning depth.

Security and guardrails are built into the inference pipeline from the start. Every MQTT payload is validated against a strict schema before it reaches the model. Model outputs pass through confidence gating and output filtering: no control action or alert fires unless the output falls within defined safety bounds. For any cloud-facing interaction, prompt injection defences, input sanitisation, and adversarial robustness checks are part of the deployment standard. Model weight updates pushed from the cloud training pipeline are cryptographically signed and verified on the gateway before acceptance. Household data never leaves the device in identifiable form.

The result is context. The inference engine knows it is mid-morning, that the kitchen has been active for twenty minutes, that movement has recently passed from the bedroom to the bathroom, and that this pattern is consistent with a normal Tuesday. That contextual model is what distinguishes a care platform that understands a household from one that merely watches it.

Room-level understanding without cameras

Spatial awareness comes from the inference stack interpreting a sparse sensor array: motion detectors, door contacts, temperature sensors, and the passive sensing capability of the Wi-Fi mesh itself. No cameras. Room-level occupancy is not read from any single signal. It is inferred from the pattern across all inputs simultaneously, updated in real time as the household moves through the day.

The gateway builds a live spatial model of the home: which zones are active, which have been quiet, and for how long. That model is the foundation for everything else. Without spatial context, an extended period of inactivity in a single room might be unremarkable. With it, the platform can assess whether that represents a normal rest pattern or something that warrants attention.

Advice, alerts, and intelligent automation

Alerts are generated by the inference engine reaching a confidence threshold, not by a sensor crossing a hardcoded value. That distinction matters. A confidence-based alert is informed by the full contextual picture: time of day, recent activity history, deviation from learned norms, and the spatial model of the home. The same sensor event that would trigger a false alarm in a naive threshold-based approach is assessed in context before anything fires. False positive rates drop significantly, and that is the single most important factor in whether care technology actually gets used over time.

Advice works the same way. Rather than surfacing every anomaly as an alert, the inference engine generates proactive recommendations grounded in what it has learned about the household: a suggestion to adjust a routine, a note that activity patterns have shifted over the past week, a prompt for a welfare check based on accumulated context rather than a single event.

Intelligent automation is the output layer. When inference reaches a defined confidence level, MQTT-driven actions execute on the gateway without waiting for a cloud response: a notification is dispatched, a contact is called, a door is unlocked for a first responder, a connected device is adjusted. The latency from trigger to action is measured in milliseconds. The cloud is not in the critical path.

The data layer that supports all of this is straightforward by design. High-frequency sensor telemetry goes into a time-series database. That is the structure built for windowed, pattern-based queries: what was happening in the kitchen an hour ago, has bathroom activity declined over the past week. Significant events and alert triggers go into an event store alongside their full inference context, creating a sequential, auditable record for care continuity and review. The cloud training pipeline ingests anonymised, aggregated event data from the deployed fleet, retrains the models, and pushes updated weights back to every gateway. What improves it is population-scale learning. Individual household data stays on-device throughout.

Why explainability matters

In assisted living, confidence matters as much as detection.

Good care technology should not just say that something happened. It should also help a carer, operator, or family member understand why an alert was triggered and how certain that assessment is.

Why this matters