AI Focus Classroom : A student-activated tutor that detects confusion and personalizes learning.
Team name
NeuralBits
Use of AI tools
Tools Used: Gemini, GitHub Copilot, and Claude 3.5 Sonnet.
Purpose & Impact:
Gemini (Legal & Logic): Researched EU AI Act and GDPR Article 22 compliance. This helped us quickly pivot from a high-risk "surveillance" model to a privacy-first model. Also used to engineer prompts for our "Socratic Mode."
GitHub Copilot (Backend): Accelerated coding for our FastAPI backend and SQLAlchemy database schemas, reducing manual boilerplate by roughly 40%.
Claude 3.5 Sonnet (Frontend): Rapidly prototyped the custom React video player and refined the UI to ensure it remained lightweight for low-bandwidth users.
External feedback & contributions
Johanna F (T-Twice team): Highlighted critical GDPR/EU AI Act risks regarding "silent monitoring" and identified our high-bandwidth infrastructure flaw.
Justine Cassell (Inria / CMU): Challenged us on academic integrity and how to prevent students from copying AI answers for tests.
Chinmay Das & Nipun Ranchhod Navadia: Pointed out that rewinding a video often means "revision," not just "confusion," requiring a UX fix.
KANKEU TAMEGHE & Tayeb Bouchikhi: Raised foundational concerns regarding data privacy and the platform's long-term educational scope.
Initial contribution
AI Shadow Classroom : Silent background monitoring that detects student confusion and to personalize learning.
Final contribution
The Problem: Students in Tier-2 and rural classrooms often fall behind in silence. Existing tools favor students with high-speed internet, premium devices, and strong English.
The Solution: NeuralBits is an AI-powered learning platform built into a custom React video player that detects when a student is stuck and provides immediate, localized support.
Core Features:
Behavior Tracking: Monitors rewind patterns and pauses to detect confusion in real-time.
Localized Support: Uses the Bhashini API to explain concepts in the student's regional language.
Automated Catch-up: Uses Whisper and LLMs to generate structured notes and personalized quizzes based on the specific moments a student struggled.
Reflection on the process
Phase 2 feedback fundamentally transformed our architecture. We realized our initial concept violated EU AI Act standards for educational profiling and required too much internet bandwidth.
To fix this, we made three major pivots:
From Surveillance to Empowerment (Privacy): We completely removed "silent monitoring." Students now use an opt-in "Focus Dashboard." Instead of auto-pausing the video, the player shows a non-intrusive prompt: "Need help, or just revising?"
Asynchronous Architecture (Infrastructure): To serve low-bandwidth users, we moved heavy processing (Whisper transcriptions, LLM summaries) to our backend. They process before the lecture is served, keeping the live video smooth on slow connections.
Socratic Mode (Academic Integrity): To prevent cheating, the AI no longer gives direct, copy-pasteable answers. It now acts as a Socratic tutor—asking guiding questions and using analogies to help students solve the problem themselves.
Share