Context
Our client works in education technology. Their question was a teaching one before it was a technical one: a teacher managing thirty students cannot watch all of them at once, and the students who quietly disengage are the ones hardest to notice and easiest to lose. The loud signals — a disruption, a question — announce themselves. The quiet ones do not.
The brief was to surface the quiet signals while the lesson is still running.
Problem
Engagement is visible in behaviour. Whether a student is reading, writing or raising a hand is observable, and a teacher standing at the front of a classroom reads those signals constantly. What no teacher can do is track thirty of them continuously while also teaching.
The binding constraint was latency, not accuracy. Analysis after the lesson is a report. Analysis during the lesson is something a teacher can act on, and only the second one changes what happens to a student that day. That meant processing video streams live, at classroom frame rates, on whatever hardware a school actually has — which is not a rack of GPUs.
The second constraint was the alert threshold. A system that fires too often teaches the teacher to ignore it, at which point it is worse than nothing because it has consumed attention and returned noise. Precision mattered more than recall here: a missed signal costs one observation, a false signal costs the teacher’s trust in the whole system.
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 over a single end-to-end model was a deliberate decision rather than a default, and it is the decision worth explaining. A large end-to-end model would likely classify better on a benchmark. It would also cost more per frame, need more hardware than a classroom has, and be considerably harder to reason about when it got something wrong. Splitting feature extraction from classification buys two things that mattered more than benchmark accuracy: inference cheap enough to run continuously on modest hardware, and a pipeline where a misclassification can be traced to a stage rather than shrugged at.
That trade-off is the whole engineering argument of this project. The model that wins on a leaderboard is rarely the model that survives a school’s IT budget.
What we built
- Real-time processing of classroom video streams at lesson frame rates
- Feature extraction from video frames
- Classification into activity categories — reading, writing, hand raising
- Live alerting to teachers based on classification output, tuned to fire rarely enough to stay worth reading
- An inference path sized for commodity classroom hardware rather than dedicated accelerators
Results
Teachers receive live signals during the lesson rather than a report afterwards, which was the outcome the engagement was scoped around.
Accuracy and latency measurements for this engagement are not published.
Related
Retrieval-Augmented Generation for an Education Platform · Serving a 72B Open-Source Model Under Concurrent Load · AI Product Engineering
