An IoT-plus-ML system that gives building operators real-time visibility into HVAC health and predicts failures before they disrupt occupants.
A US smart-buildings company wanted to give operators real-time visibility into HVAC performance across residential and commercial properties, where failures are expensive, disruptive, and usually discovered only once occupants complain.
The problem spanned the full stack: instrumenting HVAC systems, moving high-frequency sensor data reliably, and applying ML to predict issues — all robust enough to run across many buildings.
We handled IoT instrumentation and hardware design to capture HVAC telemetry reliably across building types.
A backend built for real-time time-series processing ingests and stores high-frequency sensor data at scale.
ML models predict HVAC issues before they become failures, with dashboards and alerts that give operators time to act.
Most HVAC systems in the field have no native telemetry — they're mechanical systems with, at best, a basic thermostat interface. The first layer of this project was hardware: designing retrofittable sensor modules (temperature, vibration, current draw, refrigerant pressure where accessible) that could be installed without replacing the underlying HVAC unit, and an edge gateway per building that aggregates sensor readings locally before transmitting upstream. Getting this right mattered more than any later modelling decision — noisy or missing sensor data makes every downstream model worse, so the hardware and firmware were built to handle intermittent connectivity, local buffering, and basic sanity-checking before data ever leaves the building.
HVAC telemetry is high-frequency and high-volume once you're running it across many buildings — vibration and current draw are sampled far more often than, say, a daily sales figure. The backend uses a time-series-optimised store with a real-time ingestion path, so operator dashboards reflect current state (not yesterday's batch) and the ML layer has a continuously updated feature store to read from. Data retention is tiered: high-resolution recent data for the predictive models, downsampled longer-term data for trend analysis and reporting.
The differentiator from a basic monitoring dashboard is that the ML layer is trained to predict failure modes before they happen — a compressor showing early vibration-signature drift, a refrigerant leak showing up as a slow pressure trend before it's severe enough to trip an alarm. Models are trained on labelled historical failure data where available and supplemented with anomaly-detection approaches for failure modes without enough labelled examples yet. Predictions surface in the operator dashboard as a ranked list — which units are trending toward an issue, and how much runway there likely is before it becomes occupant-facing.
Operators gained predictive visibility into HVAC health across their buildings — catching developing issues before they became failures, and turning maintenance from reactive to planned.
Tell us what you're building. We'll tell you the fastest honest path to shipping it.
Start a conversation →