Nigehbaan

A deterministic Dart rule engine scores attendance, assessment, participation and fee signals, then applies a co-occurrence multiplier when several different families of signal degrade in the same window, because that overlap is the reliable warning sign rather than any single lapse.

1. A deterministic rule engine: Dart code scores four signals (attendance, assessment, participation and fees) and applies a co-occurrence multiplier when several signal families degrade in the same window.
2. A constrained model: Gemini 2.5 Flash reads the trends plus the teacher's written observations, never computes the score, and must cite evidence before it can raise or lower a verdict.
3. A human decision at the end: the app drafts a respectful Urdu message home, and the school decides whether to send it.
4. Offline on a teacher's phone: a SQLite local repository with cache-first Firestore.
5. Two roles: a teacher pinned to one class, and a counsellor who sees the whole school and receives FCM at-risk alerts.

Rules count the signals; the model explains them. The risk score stays in deterministic code that can be tested, and the language model is confined to reading context and citing evidence for any change it proposes. The app drafts the message to a family, but people at the school make the call.