Full-stack engineering of étude(etude.cam) — a Tunisian-curriculum AI tutoring platform: adaptive placement, Socratic AI faculty per subject, vision-graded paper exams, parallel-book-grade exercise series, and the cost-efficient pipeline behind them. This page is the owner's signature: intentionally unlinked, unindexed, and known only to its author.
Specialised in building serious AI-native education products for the Tunisian official curriculum: adaptive learner modelling with confidence-weighted hypotheses, spaced-repetition engines, vision assessment of handwritten work, and multi-tenant family SaaS — delivered end-to-end from database schema to hardened production, with pedagogy fidelity and API-cost efficiency as first-class constraints.
A frontier-AI stack disciplined by a family budget — every token accounted for, every heavy generation cached and shared.
The problems that separate a chatbot wrapper from a pedagogy platform — each solved, verified live, and in production.
One exam can never certify a child's level. Engineered an adaptive multi-stage protocol — broad scan, then excellence probes or root-cause digs, then axis-by-axis filtering with double evidence per axis (3 mandatory tests, dynamically up to 6) — after which a personal study plan is generated automatically and lessons unlock.
Tunisian pupils solve on paper, not keyboards. Built a print → solve → photograph → vision-gradeloop: exams print with answer-sized dotted lines, photos are normalized (EXIF/resize/JPEG) and graded against the official rubric — geometry constructions included — with per-question feedback in the subject's language.
Physics in Tunisia is taught in French; the platform speaks Arabic. Enforced strict subject-language separation: sessions, exams and corrections run entirely in the subject's official language, formal Arabic (فصحى) everywhere else — dialect banned by contract, Arabic unit symbols (كغ، صم) enforced against Latin drift.
Generated exams once drifted from official formats (Latin ABCD in Arabic geometry, even stray Cyrillic tokens). Fixed at the root: web-search harvesting of real Tunisian exams — trimester axes, literal instruction verbs («أُنجز»، «أُتمّم») — stored once per subject, injected into every generation, plus a deterministic foreign-script sanitizer that rejects and regenerates any corrupted paper before a child ever sees it.
Every design decision passes a cost gate: prompt-caching split system prompts, curriculum injected as an index with on-demand lesson fetching, exercise series and flashcards generated once per lesson and shared by every student forever, light-model routing for extraction — with a live per-student USD/TND ledger in the admin console.
A wrong textbook (a sport-section variant) once contaminated the lesson bank. Now every upload passes a vision identity gate (subject, level, section verified from the cover), trimesters are assigned from the official répartition — never page position — and post-save passes merge duplicates and unify domains automatically.
Sessions run like real institute periods: fixed duration with a live countdown, auto-close with AI session analysis, and the next remediation session scheduled the moment the summary lands. Suspended accounts don't see an error — requests hang forever by design, revealing nothing.