← All work
Solo builder·2024 — Present

AI IELTS writing corrector calibrated by examiner team with 7,000+ real corrections.

Real corrections
7,000+
Scoring dims
4 (TR/CC/LR/GRA)
Model
Examiner-calibrated
Output
PDF study report

Problem

IELTS writing prep is high-friction. Generic AI correctors return black-box band scores plus a "model essay" whose ideas aren't the student's — students can't reuse any of it in the real exam. Worse, 80% of weak essays fail at the outline stage, but most tools only inspect language.

What I built

  • Four-dimension band-descriptor scoring (TR / CC / LR / GRA) calibrated by an IELTS examiner team and refined with 7,000+ real corrections.
  • Outline-first diagnostic loop: before the student writes, the AI diagnoses the outline for relevance, logic, and coherence — fix the argument before the language.
  • Sentence-level rewrite engine: every sentence is tagged (grammar / lexical / cohesion) and shown in an original → improved comparison.
  • Cohesive-device analysis + auto-generated vocabulary recall and translation drills targeting each student's weak points.
  • One-click PDF study report so sessions survive the online interface.
  • Stripe-based per-use credit system with a free-trial tier tuned for first-time users.

Stack notes

Next.js App Router for the marketing and dashboard surfaces; PostgreSQL for credit ledger and session storage; a custom pipeline that composes multiple AI passes (outline → sentence-level → cohesion → vocabulary) instead of a single giant prompt — this is what lets deductions map to Band Descriptors.

Next.jsTypeScriptAIStripePostgreSQL

What I learned

  • Fixing the outline before the language lifts TR scores more sustainably than any sentence-level edit.
  • Rewriting along the user's own reasoning beats showing a perfect model essay — the goal is what they can reproduce in the exam, not what looks impressive.
  • A downloadable PDF report at the end of a correction doubles retention — users come back to re-study, not re-correct.