# OESED

> We designed a system that supports a teacher's classroom observation with signals that can be followed over time and explained. Video is processed on the device in the classroom; only numeric signals leave it.

URL: https://www.oneeyesystems.com/en/work/oesed/
Türkçe sürüm: https://www.oneeyesystems.com/tr/calismalar/oesed/index.md

- Domain: Education · support for classroom observation
- Status: Our own product

## Problem

In a crowded classroom, a teacher cannot follow how every student changes over the day, and a school counsellor often decides what to prioritise by intuition. The idea of camera-based observation raises legitimate concerns: stored footage, data attached to the wrong student, automatic sanctions, emotion prediction. What matters is pointing to changes that deserve attention, without deciding for the teacher and with privacy protected from the start.

## Approach

1. Video stays on the device. The device in the classroom detects people, tracks them through the lesson and produces explainable signals from 12 body keypoints. Raw video and raw pose data are not stored permanently.
2. Only numbers leave. Pseudonymised numeric packets are sent over an encrypted connection; no video, face crops or face embeddings leave the device.
3. Separate dimensions, personal baselines. Movement, body orientation and changes in posture are tracked as separate dimensions against each student's own baseline. No single “engagement score” is produced.
4. A daily briefing. The counsellor sees at most three alerts to review, each with an explanation of the signal. Parents see only their own child's weekly summary and their data rights requests.
5. Separated responsibilities. Video processing, identity matching, signal calculation and transfer are kept apart; analysis services have restricted access to student names.

## Critical decisions

### Uncertain records are not attributed

If a match is not certain, the data is excluded. Losing data is preferred to writing one student's data to another.

### No automatic sanctions

The system makes no educational decisions, takes no disciplinary action, performs no emotion analysis and runs no school-wide face search. The teacher has the final say.

### Event-based face verification, with separate consent

Not continuous: it runs only on events such as the start of a lesson, a new entry or a tracking conflict. It is not enabled without separate biometric consent and independent testing.

### Processing stops when consent is withdrawn

When a preference changes, processing stops on the device; a time-limited authorisation policy prevents it from continuing even when the connection is down.

### Small groups are not reported

Groups of fewer than five are suppressed in reports, so a report that looks anonymous cannot point to a single student.

### Evidence first, classroom second

Classroom use does not start until performance, matching and data leakage tests have passed in the lab.

## Details

### Architecture

| Layer | Role | Boundary |
| --- | --- | --- |
| Classroom device | Person detection, in-lesson tracking and explainable signals from 12 body keypoints | Raw video and raw pose data are not stored permanently |
| Secure transfer | Sends pseudonymised numeric packets over an encrypted connection | No video, face crops or face embeddings are transferred |
| Analysis and identity vault | Time series, personal baselines and the daily briefing; a separate identity resolution zone | Analysis services have restricted access to student names |
| Teacher and parent | Signals and explanations for teachers; parents see only their own child's summary | Human judgement comes first; no automatic sanctions |

### Scope and limits

- Prioritisation for the school counsellor: at most three alerts a day
- Clear, limited access for parents, including data rights requests
- Out of scope: automatic educational decisions, discipline, emotion analysis and school-wide face search

### Rollout sequence

| Stage | Gate |
| --- | --- |
| Finalise the architecture | One reference version with identity, session, consent and data contracts |
| Lab validation | Performance, matching and data leakage checks pass |
| Controlled classroom pilot | Approvals, role assignments and operating procedures are in place |
| Biometric evaluation | Independent acceptance decision; if it fails, biometrics stay off |

Success is judged on results measured under the same pilot conditions: device performance, matching errors reported together with data loss, agreement between signals and observer assessment, and consent withdrawal and deletion scenarios. Targets are kept separate from measured results.

---

One Eye Systems — Enterprise AI consulting and implementation · https://www.oneeyesystems.com/ · info@oneeyesystems.com · https://www.linkedin.com/company/oneeyesystems/