AI features

AI is a teaching assistant, never the final examiner. This page explains what the AI does, how it stays trustworthy, and how its cost is controlled.

What the AI does

Feature	What it produces	Where it runs
Writing grading	Band per IELTS criterion (Task, Coherence, Lexis, Grammar) + in-essay feedback + an AI-writing/integrity signal	Cloudflare Workers AI
Speaking feedback	Transcription + fluency/criteria estimate + improvement points	Workers AI (transcription + grading)
Pronunciation (read-aloud only)	Real per-phoneme accuracy, fluency, completeness	Azure Speech (optional, capped)
AI Tutor	Chat that explains grammar/vocab/answers, personalized to the student's band and common errors	Workers AI
Test extraction (OCR)	Turns an image/PDF of a test into a structured draft	PaddleOCR container + Workers AI structuring
Teaching insights	Per-attempt suggestions for the teacher, from a compact stat sheet	Workers AI

The product rule on model names

Users never see model names. Every AI surface says "AI" — never "Whisper", "Gemma", "Azure", etc. This is a deliberate product decision and applies to all user-facing copy, including manager reports.

The trust guarantee: AI bands are never final

This is the most important thing to understand about the platform.

graph LR
  S[Student submits Writing/Speaking] --> AI[AI grades]
  AI --> P["Status: provisional<br/>(shown as a suggestion, not a score)"]
  P --> Q[Teacher review queue]
  Q --> T{Teacher}
  T -- edits + approves --> F["Status: approved<br/>(now a real band)"]
  T -- low AI confidence --> Q

• Reading & Listening are objective → auto-graded → final immediately.
• Writing & Speaking are subjective → the AI band is provisional and clearly labeled as such → it becomes a real band only when a teacher approves it.
• If the AI is unsure (low confidence, or it failed), the essay is routed to the teacher's queue rather than being stamped with a fake band.
• Manager-facing surfaces (reports, profiles, dashboards) must not present a provisional band as a final score.

Why it matters for the product: it means the center can safely put AI in front of students without risking a wrong "official" score. The teacher stays in control; the AI just does the first pass and saves time.

How Writing grading stays consistent

The Writing grader is a calibrated pipeline, not a single AI guess:

1. Each criterion is scored independently; the AI is forbidden to emit an overall band.
2. The overall band is computed in code from the four criteria (with Task 2 weighted double for combined assignments) — not taken from the model.
3. Each criterion is graded three times and the median taken, to smooth out run-to-run variance.
4. Confidence is derived from how much the passes disagree and whether the AI's quoted evidence actually appears in the essay; low confidence → teacher review.
5. In-essay highlights are validated word-for-word against the student's text; anything the AI invents is dropped.

Budgets & quotas

The center pays a flat monthly budget. To stop one student (or a runaway loop) from draining it for everyone, usage is metered and hard-capped before the paid call, and the caps fail closed (if usage can't be confirmed, the call is skipped, not run).

Feature	Monthly cap	What happens at the cap
Writing gradings	1,000	New essays go to the teacher queue instead of AI grading
Speaking feedback	6,000	Falls back to the built-in heuristic score
AI Tutor turns	6,000	Returns a "monthly limit reached" message
Pronunciation (Azure)	manager-set minutes	Falls back to the in-app recognition estimate

Fixed in 2026

Writing grading previously counted its 3 internal median passes as 3 gradings, burning the 1,000/month budget ~3× too fast and throttling teachers at a third of the intended volume. It now counts one grading = one unit, so metering and the quota agree.

The manager watches all of this on the Reports → AI usage panel (labeled "Sử dụng AI"), which shows monthly usage per feature against the caps — so the budget is visible before it's exceeded.

Honest limits

• Pronunciation is only measured on the read-aloud tool (known target text). In open-ended Part 2/3 speaking, pronunciation and intonation are band-derived estimates, clearly labeled as such — the transcription model is accent-robust and can't reliably measure mispronunciation.
• Absolute AI band accuracy is unvalidated against human labels, which is exactly why teacher approval is mandatory. The AI is consistent, not certified.
• AI never reads test content in OCR extraction — real OCR extracts the text faithfully; the AI only structures it into a draft the teacher reviews.