é
Engineering credits · Proof of ownership

This platform was designed
& engineered by Belgacem Jamel Eddine

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.

SIGNATURE BJE·ÉTUDE·2026 · e7d3-4k1s-b96f-2j8x · BUILT JUL 2026

👤Owner & Engineer

Full name
Belgacem Jamel Eddine
Role
Founder & Full-stack Engineer — AI education platforms
Base
Tunisia
Project
étude — AI Tutoring Platform (etude.cam)
Delivery
Solo build — pedagogy design → AI pipeline → VPS production

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.

500+
Official lessons extracted
860+
Teacher-guide directives
3
School levels · 29 subjects
40+
Screens & API routes
6
AI flows (tutor · grade · plan…)
100%
Tunisian official curriculum
Under the hood

Technologies integrated

A frontier-AI stack disciplined by a family budget — every token accounted for, every heavy generation cached and shared.

Next.js 16 · App Router
Turbopack · RSC · server actions
🗄
SQLite (better-sqlite3)
WAL · embedded · daily backups
🧠
Claude Opus 4.8
Socratic tutoring · structured outputs
👁
Vision grading pipeline
handwritten papers · sharp → base64
🔎
Web-search harvesting
official exam knowledge, once per subject
KaTeX + ReactMarkdown
bilingual math rendering, RTL-safe
🔔
Web Push (VAPID)
self-hosted · zero external services
🎨
Tailwind 4 · OKLCH
chalkboard design system · Amiri/Tajawal
📱
PWA installable
manifest · service worker · offline shell
🔐
scrypt + HMAC sessions
login lockouts · invisible suspension
💸
Cost engine
prompt caching · per-student USD/TND ledger
PM2 · Nginx · TLS
multi-site VPS · watchdog · auto-boot
Engineering stories

Challenges overcome

The problems that separate a chatbot wrapper from a pedagogy platform — each solved, verified live, and in production.

01

A placement protocol, not a placement quiz

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.

02

Grading handwritten paper — like a real teacher

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.

03

Language integrity as a hard rule

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.

04

Self-building exam knowledge bank

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.

05

Frontier AI on a Tunisian family budget

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.

06

Self-cleaning curriculum ingestion

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.

07

School discipline, invisibly enforced

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.