Comfort is six things, not a thermostat.
Air temperature alone is a poor proxy for how a room feels. Five of the six are measurable with sensors that cost less than the switch they sit behind.
What people feel is operative temperature — air temperature combined with the radiant temperature of the surrounding surfaces, air movement and humidity. A room at 24 °C with a west wall that has been in the sun since two o’clock does not feel like 24 °C.
Controlled through the air conditioner, fan speed and shading together rather than the compressor alone.
Sensing — SHT45 for air, MLX90640 thermal array for radiant surfaces.
The axis nobody measures and everybody feels. Carbon dioxide above roughly 1000 ppm is the single largest cause of the sensation people describe as stuffy, and of poor sleep in closed bedrooms. It is invisible, cheap to measure properly, and almost universally ignored.
Drives fresh-air ventilation and purification, and stops the house sealing itself shut for the sake of the cooling load.
Sensing — true NDIR CO₂, PM2.5 and VOC. Not the estimated-CO₂ sensors that infer from VOC.
Independent of temperature and often the real complaint. Hyderabad sits at 70–85 % through the monsoon; the comfort band is 40–60 %, and dry mode alone rarely gets there.
Two jobs at once: seeing, and keeping a circadian rhythm intact. Needs illuminance and spectrum control together, 1800–6500 K, with blinds acting as the daylight actuator before any lamp comes on.
Mostly a passive property of how the building was made, but measurable. Sound pressure level only — no audio is recorded, buffered or transmitted. Used to avoid running noisy loads during sleep windows.
The enabler for all of the above — there is no point conditioning a room nobody is in. Millimetre-wave sensing detects a person sitting still, which infrared cannot, and resolves zones within a room rather than treating it as one box.
It also produces no image. That is deliberate: a building that measures presence should not become a building that watches people.
The overrides are the training data.
Every time someone reaches for a switch, they are correcting the system. That correction is more honest than any preference screen.
Comfort is personal and it drifts — same room, same numbers, different day. So the model’s job is not to hit a published standard. It is to learn the band each household actually lives in, per room and per time of day, and to notice when that band shifts.
What it proposes still passes through the same broker as everything else. It can suggest a warmer evening setpoint; it cannot quietly apply one. If the model is unreachable, the house falls back to a deterministic schedule rather than getting confused.
No cameras and no audio. Presence comes from millimetre-wave sensing, which detects a body without producing an image. The acoustic axis measures level only.
State stays local — the property model, event history and decisions live on the core in the house. Cloud services exist for model calls, remote access and backup, and are optional rather than load-bearing.
The audit trail exists so behaviour can be explained and disputed by the owner. It is not a feature we monetise and it is not shipped anywhere by default.