ALTF4

Inspiration

Every EdTech platform rewards correct answers, not actual learning. We watched students ace quizzes by memorizing, copying from AI, or guessing confidently — then forget everything a week later. The question that haunted us: what if a platform could tell the difference? That's EALE.

## What it does

EALE measures Durable Understanding — not just whether you answered correctly today, but whether you'll remember it next week, apply it to new problems, and actually know what you know. It computes a Durable Understanding Score (DUS) across four dimensions: Mastery, Retention, Transfer, and Calibration. A Chrome extension embeds micro-assessments into any webpage — YouTube, Notion, Canvas. It generates AI video lessons via OpenAI Sora, monitors attention with YOLOv8, and flags students using AI to cheat through our AI-Dependency Fingerprinting Algorithm.

## How I built it

• Backend: FastAPI + PostgreSQL + APScheduler for spaced repetition scheduling
• Frontend: Next.js 14 + TanStack Query for real-time student and faculty dashboards
• Chrome Extension: MV3 with Shadow DOM overlay — works on any site without breaking page styles
• LLM Pipeline: GPT-4o for question generation, grading, and narration scripts; Sora for 1280×720 video lessons; TTS-1-HD for audio
• Computer Vision: YOLOv8 (via FastAPI microservice) for real-time attention monitoring via webcam

## Challenges I ran into

• Getting the Chrome extension to work on HTTPS pages (YouTube, Canvas) — solved via background service worker message relay to bypass CSP restrictions
• Sora video generation takes 60–90 seconds — built a polling system with graceful HTML canvas fallback
• Designing the AI-dependency fingerprint without false positives — required a 3-signal composite score (retention collapse + transfer gap + calibration paradox)
• The youtube-transcript-api v0.6.2 was silently broken — tracked down a ParseError and pinned to v1.2.4

## Accomplishments that I'm proud of

• The DUS formula actually works — Alice scores 47, Bob scores 18, a genuine learner scores 82, and the numbers feel right
• AI-dependency fingerprinting catches the cheating signature that accuracy alone can never see
• The extension overlays cleanly on YouTube, Notion, and Canvas with zero style bleed thanks to Shadow DOM
• End-to-end in one hackathon: video generation, attention monitoring, spaced repetition, calibration scoring, faculty analytics — all wired together

## What I learned

That measuring learning is a fundamentally harder problem than measuring performance. Ebbinghaus was right in 1885 and still right today — forgetting is the default, not the exception. Building a system that fights forgetting by design rather than by accident taught us more about pedagogy than any course we've taken.

## What's next for ALTF4

• LMS integrations (Canvas, Moodle, Google Classroom) via LTI 1.3
• Longitudinal DUS tracking with learning velocity graphs
• Peer cohort benchmarking — see how your retention compares to classmates
• Multi-modal assessments: diagram labeling, code execution, audio responses
• Mobile companion app with push-based spaced repetition reminders

Built With

• computervision
• docker
• fastapi
• javascript
• python
• sqlite
• yolov8