Skip to content
All work
IEEE TALE 2026 · acceptedML SystemsJan 2025 - Apr 2025Project Lead

AffectLearn

Sentiment-aware adaptive STEM tutor

View code
learner approval
85.2%learner approval
BLEU
0.451BLEU
ROUGE-L
0.772ROUGE-L
Flesch readability
71.1Flesch readability

The problem

Most AI tutors answer every student the same way. They do not notice when a learner is confused, frustrated or bored - so the teaching never changes when it should.

What I built

AffectLearn listens to text, voice and face: RoBERTa reads sentiment, Whisper transcribes speech, a CNN reads facial cues - then a LoRA-tuned TinyLLaMA adapts its tone, difficulty and depth.

Highlights

  • LoRA-tuned TinyLLaMA - runs on plain consumer GPUs
  • RAG grounded in FAISS-indexed OpenStax textbooks - answers stay accurate
  • Session feedback stored in Supabase for a picture of the learner over time

How it works

  1. 01Learner inputText · voice · webcam
  2. 02Multimodal readRoBERTa · Whisper · CNN
  3. 03RAG retrievalFAISS over OpenStax
  4. 04LoRA TinyLLaMA
  5. 05Adapt tone & difficulty
  6. 06Session logStored in Supabase
Next projectEcoSphere-AIExplainable AI for climate-smart housing