Model development · Stage 2

Learning from the kitchen, one reading at a time.

The first models are now publishing predictions, but they are still very weak because they need more training data. Scentinel continues gathering sensor readings and human votes to improve them.

Live inference

Latest prediction

Loading the latest model output…

Checking for a prediction

Kitchen activity—Loading
Food prediction—Loading
Cooking method—Loading
Development path

From observations to useful predictions

01 Complete
Gather initial data

Initial gas fingerprints and human votes enabled the first models. New labels continue to improve them.

02 Active
Improve models and run inference

Train and validate improved models while publishing kitchen activity, food predictions, and confidence from the best available model.

Why votes matter

Sensor readings need human context.

A gas fingerprint can show that the air changed, but it cannot initially explain why. Votes connect a sensor window to known kitchen activity, food, and cooking methods, creating labels that the model can learn from.

Contribute a label
Live outputs

What the models report

Kitchen activity
Estimated likelihood that cooking activity is occurring.
Food prediction
The food class most consistent with the recent gas fingerprint, when a food class is available.
Cooking method
The predicted cooking method, when the model provides one.