Supportive AI ecosystem for Malaysian small-class tuition centres
Challenge: Malaysian tuition centres lost UPSR. The new PBD + UASA continuous assessment is far richer — but the tools centres have (Google Sheets, WhatsApp, paper) cannot synthesise that signal into early supportive action. By the time end-of-term reports surface a problem, three months are lost and the relationship between child, teacher, and parent has eroded.
Solution: North Star Learning Center is a three-role AI ecosystem. The centre manager sees critical escalations and AI-drafted parent communications. Each specialist teacher sees their class with per-child support suggestions and a Help-a-Colleague nudge when peer mentorship would help. Each parent receives a warm monthly story written by Gemini — never raw scores, only growth and rationale.
Manager Home auto-surfaces children needing attention. Click through and the AI proposes a specific support intervention — keep the current specialist, try a new approach, monitor in two cycles — with reasoning.
Supportive signals only — workload, momentum, retention risk. Never a leaderboard. Teachers are protected from being ranked or scored.
The teacher's home view. Per-child 2-line support suggestions grounded in real dev notes. Plus a "Celebrate This" panel surfacing wins to acknowledge in class.
For monthly Science lab sessions, the AI suggests 3 groups of 3-4 children based on past collaboration history. No child is ever placed alone. Includes a "Why this combo" reasoning line for each group.
When a junior teacher's class shows gaps, the AI flags it on the senior teacher's screen and drafts a 2-sentence peer-mentor note in supportive Malaysian English. No ed-tech we know of does this.
Gemini drafts a warm monthly letter for every family. The manager reviews and edits before sending. Letters never contain raw scores — only growth, rationale, and one concrete way the family can help at home.
Manager-only week-at-a-glance: capacity (76 slots open by subject), 185 children, 22 specialists, 44 classes with a 10-child cap. One screen, no clutter.
On the parent view, a plain-English explanation of why their child was assigned to a particular specialist. Transparency without surveillance.
For every featured family, Gemini drafts a warm 4-6 paragraph monthly story grounded in the child's recent dev-note observations. Hard constraints: never include numeric scores, never compare to other children, always mention one growth moment and one current focus area. Runs in ~2.5 seconds at temperature 0.4.
When the AI suggests a senior teacher reach out to a junior colleague, Gemini drafts a 2-sentence supportive note in Malaysian English. Constraints: exactly 2 sentences, peer-tone (not lecturing), first name only, no formal sign-off. Runs in ~1.1 seconds at temperature 0.6.
Eight named, typed, deterministic tools handle all SQL queries and scoring logic — centre_snapshot, at_risk_children, teacher_wellbeing, teacher_home, centre_directory, student_profile, parent_story, draft_peer_note. Clean architectural split between deterministic math and warm narrative.
Wellbeing, discipline, cleanliness, and collaboration are NEVER scored. Children never access the platform directly. Teacher wellbeing is a supportive signal, never a leaderboard. These constraints are encoded in the data model itself — not a layer that can be turned off.
gcloud run deploy --source . command/api/manager/home, /api/teacher/{tid}/home, /api/parent/{sid}/story — each role gets a tailored payload, not a god-object dump.python prepare_data.py — One-time data migration (renames students, seeds demo rows, enforces 10-child class cap). Idempotent.python main.py — Start the FastAPI server on port 8080start.bat — Windows one-click launcher (Python + venv + browser open)Dockerfile and deploys in ~3 minutes