Publish this one carefully. Computer vision applied to children in classrooms is a live ethical debate, and a page that describes it purely as a technical achievement will read badly to some of the education buyers you want. The version below leads with what the system does and does not do. Recommend keeping the client anonymous and adding the privacy paragraph once the facts are confirmed.
Context
Our client works in education technology. A teacher managing thirty students cannot watch all of them, and the students who quietly disengage are the ones hardest to notice and easiest to lose.
Problem
Engagement is visible in behaviour — whether a student is reading, writing, or raising a hand — but no teacher can track thirty of those signals continuously while also teaching.
The technical constraint is that this has to run live. Analysis after the lesson is a report; analysis during the lesson is something a teacher can act on. That means processing video streams in real time, at classroom frame rates, accurately enough that alerts are worth reading rather than noise a teacher learns to ignore.
Approach
The pipeline is feature extraction followed by classification. Video frames are processed to extract features, and those features pass through machine learning models — convolutional neural networks and support vector machines — to classify activity into recognisable behaviours.
Choosing CNNs and SVMs rather than a single end-to-end model is a real engineering decision, and worth explaining: it typically buys interpretability and lower inference cost against a large model, which matters when the thing has to run continuously on modest hardware in a school.
NEEDS INPUT: whether that was the reasoning, what the inference target was (frames per second, on what hardware), and how the models were trained and validated.
What we built
- Real-time processing of classroom video streams
- Feature extraction from video frames
- Classification into activity categories — reading, writing, hand raising
- Live alerting to teachers based on classification output
Privacy and scope
NEEDS INPUT — and this section should not ship without it. Buyers will ask, and the honest answers are a selling point if the system was built responsibly:
- Is footage stored, or processed and discarded?
- Is the output aggregate class-level engagement, or per-student identification?
- Is any facial recognition or individual identification involved?
- What consent model was in place?
Results
NEEDS INPUT. Classification accuracy on a held-out set, latency per frame, or classroom sessions processed.
Related
Retrieval-Augmented Generation for an Education Platform · Serving a 72B Open-Source Model Under Concurrent Load · AI Engineering