NeuroAdapt
A learning platform that notices when a student is struggling and changes the lesson, with no diagnosis required.
The question
Can a platform tell that a student is struggling from how they behave, without asking them to disclose anything?
NeuroAdapt watches a handful of behavioural signals and changes the lesson when they point to trouble.
Why it exists
About 20% of undergraduates are neurodivergent, and the project’s documentation puts their course completion 26% lower, while the platforms they use are built for an average student. NeuroAdapt adapts without asking for a diagnosis or an accommodation form.
How it works
From behaviour to a changed lesson
- 1
Observe
The browser summarises dwell, jitter, tab switching, stalls and the learner’s format choice into five numbers, every 30 seconds.
- 2
Decide
The policy scores the six actions. Below 60% confidence it holds course.
- 3
Generate
The chosen action is turned into new material.
- Text
- Llama 3 on Groq, with Ollama as a fallback.
- Images
- Stable Diffusion, with a drawn SVG as a fallback.
- Voice
- Kokoro text to speech.
- 4
Feedback
The outcome and a reward are stored, and the policy is retrained offline from them.
Decisions
Only a five-number summary leaves the browser
The raw signals stay on the learner’s machine. The server receives the summary vector and nothing else.
A confidence gate on every intervention
The policy acts only when it is at least 60% sure. Below that the lesson is left alone, so the system interrupts less and when it does, it is more certain.
A stability-first reward
The reward penalises overload and a drained energy bar and gives a bonus for stability, so the policy is not rewarded for stimulating the learner.
A classical model sits next to the quantum one
The same code base has a classical Double DQN to compare against, and a rule-based policy takes over when no trained checkpoint is available.
Evidence
- Built
The browser observer, the backend, the quantum policy, the retraining script, the generation engine and a Docker setup are all in the repository.
Not built yet
- A trained policy checkpoint (not included in the repository)
- Online learning: the policy is retrained offline
Built with
Frameworks and libraries
- Next.js
- FastAPI
AI and models
- PennyLane
- LLaMA-3
- Stable Diffusion
Infrastructure and services
- Docker
Methods
- Double DQN