BoilGuard

checking service…

A range-hood camera that sees a pot about to boil over, predicts when with a physics-informed model, and turns the heat down before it spills. Built on OpenCV 5 and AWS.

Choose a scenario and press Turn on the stove.
You'll get the camera view annotated by OpenCV, the forecast and every agent decision.

Vision pipeline (OpenCV 5)

Find the rim. Median of the first frames, CLAHE, Canny, findContours + fitEllipse, scored by edge support. Refined to the inner edge of the bright lip by a per-angle sub-pixel fit.
Unwrap. remap into an elliptical polar image, so a rising surface is a front moving outward.
Stabilise. Rim-anchored phaseCorrelate with a Hanning window cancels hood vibration; hands crossing the rim are rejected.
Normalise light and steam. Trimmed per-angle regression on the rim lip gives gain/offset; low gain = steam.
Find the foam front. The inner wall is static background: a vectorised change-point search per angle, median over angles.
Depth. Pinhole geometry: d = H (r_rim / r_surface − 1).
Agitation. Farneback optical flow flags boiling onset.

Prediction and action

Physics. Foam grows exponentially then saturates. cv2.KalmanFilter tracks log-foam, growth rate and its change.
Uncertainty. Monte-Carlo samples of the posterior give P(overflow within the hob's heat lag).
Act. Alert, then lower heat via AWS IoT Core before the lag runs out (with cancel window / approval modes).
Verify. Watch that foam growth reverses; if not, burner off + alarm.
See better. Steam hides the rim → switch the hood fan on.
Explain. Amazon Bedrock (Nova) writes the notification and incident report from the numbers only.

BoilGuard

Kitchen scenario

Agent decisions perception → decision → action

Incident report

Scenario

Your own clip

Held-out test set

Vision pipeline (OpenCV 5)

Prediction and action