The problem LMS by TLE Terminator solves

The problem TLE Terminator: The Student Success Engine solves
LMS by TLE Terminator bridges the "Implementation Gap" in modern education—the disconnect between teaching content and students actually understanding it. It solves critical inefficiencies in online and resource-constrained learning environments:

Online teachers often feel like they are "teaching to the void" with no visibility into student engagement. Our Attention Engine acts as a digital invigilator, providing real-time feedback on student focus (gaze/head pose), enabling teachers to pivot instantly when the class zones out.

Millions of students in underfunded or rural schools learn science theoretically because they lack physical laboratories. Our Virtual STEM Ecosystem brings Physics, Chemistry, and CS labs directly into the browser, allowing students to experiment safely and practically without expensive infrastructure.

3.Eliminates "Revision Burnout" & Repetitive Doubts Students waste hours re-watching lectures for notes, and teachers answer the same doubts repeatedly. Our AI Layer solves this by:

Auto-Summarizing
Lectures: Generating instant, structured revision notes from audio.
Context-Aware
Tutoring: Handling routine doubts instantly (chatbot support)
4.Provides "360° Student Guidance"
Most platforms stop at academic content. We solve the "Directionless Student" problem by integrating Personalized Career Guidance, ensuring students don't just learn what to study, but understand why it matters for their future.

Challenges I ran into

Challenges we ran into
Challenges We ran into:
🔴 Problem
Initial WebRTC peer-to-peer (mesh) setup only supported 1-to-1 communication and did not scale.
High bandwidth and CPU usage with multiple participants.
Complex signaling logic and frequent sync & connection drops.
tldraw whiteboard updates were not syncing correctly, causing latency and inconsistent board states.
During deploying to production, faced issues with license limits for tldraw.
🛠 Solution
Migrated to Stream.io SFU architecture for scalable multi-user video streaming.
Optimized whiteboard sync using incremental event updates, batching, and throttling.
Reduced latency and ensured real-time board consistency across users.
Made temporary license arrangements for tldraw ,planning to build a custom whiteboard or move to open source solution in the future.
2. PDF Upload & Lecture Notes System
🔴 Problem
Broken download links, duplicate file overwrites, and inability to directly download PDFs from Cloudinary.
🛠 Solution
Implemented Multer-based upload pipeline with validation.
Used timestamp + UUID naming for uniqueness.
Uploaded PDFs in Cloudinary raw format for direct downloads.
3. Video → Audio Differentiation (Cloudinary Streams)
🔴 Problem
Incorrect audio extraction, multi-channel mismatches, transformation failures, and high streaming latency.
🛠 Solution
Used explicit Cloudinary audio transformations with fixed codecs and sampling rates.
Added format validation, retry logic, and optimized delivery endpoints.
Built a stable pipeline for AI transcription & summarization.
4. AI-Based Lecture Summarizer
🔴 Problem
Converting long lecture videos into accurate transcripts.
Handling large audio files without blocking the main application.
Preventing system slowdowns during peak uploads.
🛠 Solution
Implemented OpenAI Whisper for high-accuracy speech-to-text transcription.
Designed an asynchronous queue-based processing pipeline to handle heavy transcription jobs without blocking the server.
Used pooling and rate-limiting to manage concurrent requests and maintain system responsiveness.
5. Attention Engine
🔴 Problem
Model Accuracy vs Speed Trade-off:
High-accuracy models reduced speed, while lightweight models reduced reliability. Finding the right balance for hackathon constraints was tough.

Inconsistent Lighting Conditions
: Different lighting environments (low light, backlight, shadows) affected face detection accuracy, leading to false attention drops.

Real-Time Webcam Performance:
Processing live webcam frames caused lag and high CPU usage, especially on low-end devices. Optimizing frame rate without losing accuracy was challenging.

🛠 Solution
Reduced frame rate (processed every nth frame) and resized frames before processing and moved heavy computation to backend and used lightweight CV pipelines
Used lightweight pretrained models optimized for real-time inference and applied threshold-based attention scoring instead of complex deep models
Balanced accuracy by combining multiple simple signals (face presence + head pose)