The life of one thermal frame
The honest way to explain how Stovyn handles your data is to follow one piece of it. This is the biography of a single infrared frame — captured, calibrated, judged, and erased on your countertop in about sixteen milliseconds — and of the rare event that outlives it.
WHAT THE DEVICE SEES
768 temperatures, several times a second.
Over a hundred million readings a day — analyzed and erased on your counter, never uploaded.
WHAT OUR CLOUD SEES
{
"event": "unattended_risk",
"burner": 2,
"zone_max": "214 °C",
"sustained": "5m 12s",
"presence": false,
"ack": null
}One small event — only when something’s wrong.
About the size of a text message. A conclusion, not a recording.
- t + 0 ms
A frame is born
Frame #38,412 of the afternoon: the 32 × 24 thermopile array captures the long-wave infrared your cooktop emits — 768 radiometric pixels, each an independent temperature estimate. No lens pointed at you, no light needed; a thermopile reads heat, so it sees a simmering pan identically at noon and at midnight.
- t + 4 ms
It learns what it’s looking at
Radiometric normalization first: per-pixel gain and offset correction, then ambient-temperature compensation — so a hot August kitchen doesn’t masquerade as a hot burner. Then spatial segmentation: the frame is partitioned into per-burner zones calibrated to your stove’s layout during setup. The frame stops being a picture and becomes five numbers: one peak temperature per zone.
- t + 9 ms
It gets a memory check
A single frame can lie — a steam plume drifts by, a hand reaches across. So each zone’s reading joins a rolling window: exponential smoothing plus median filtering across recent frames. Transients wash out; only sustained heat survives the filter. This is the step that separates “someone is cooking” from “something is wrong.”
- t + 15 ms
It is judged
The filtered values drive a finite-state machine per burner — idle → cooking → unattended-risk → local-alarm → escalation — with hysteresis thresholds (different bars to enter and leave a state, so it never flaps at the boundary) scaled by your sensitivity setting. Today, burner 2 is at a rolling boil with you standing right there: state stays cooking. Nothing is wrong.
- t + 16 ms
And then it dies
Its verdict delivered, frame #38,412 is overwritten in the device’s memory. It was never written to disk, never compressed, never uploaded. On a normal day, this is how every single frame ends — a hundred-million-plus readings live and die on your countertop. Your kitchen’s rhythm — when you cook, what you cook, who wandered past — stays in your kitchen.
- the exception
When risk sustains, an event is born instead
Minutes later: the pot still boils, presence hasn’t been detected, the state machine crosses into unattended-risk. Now — and only now — something leaves the device: a structured event record over TLS 1.2+, roughly the size of a text message. On the Pro, the camera wakes for one still image; a cloud vision model reduces it to a detection result (pot? person? flame? smoke?) and the image is not retained. The audio pipeline works the same way — band-power analysis on the device recognizes a kettle whistle; no recording is ever made.
- t + 2 s
The event outlives the frame
In the cloud (SOC 2 Type II infrastructure, AES-256 at rest), the event opens an acknowledgment window: your phone gets the push while the device sounds its ~85 dB alarm at the stove. One tap resets everything. If the window lapses, escalation proceeds — SMS to up to five trusted contacts, each of whom gave explicit consent before a single message could ever be sent. Every step is idempotent and logged: a network retry can’t double-page your daughter.
The receipts: data minimization at a glance
| Signal | Processed | What we keep | Guarantee |
|---|---|---|---|
| Thermal frames (768 px, several Hz) | On-device | Zone summaries + events only | Raw frames are never uploaded |
| Audio (band-power detection) | On-device | Detection events only | Recordings are never made or stored |
| Camera (Pro, event-triggered) | Single still → cloud vision model | Structured detection result | Images are not retained; camera can be disabled |
| Account & contacts | Cloud | Email, consented contact numbers | No sale of personal data |
The binding version of these commitments lives in our Privacy & Data page.
We use cookies to improve your experience
We use essential cookies to make our site work. With your consent, we may also use non-essential cookies to improve user experience. Learn more
