Learn.AI

Inspiration

The world, and especially tech, is advancing rapidly, but modern education hasn’t kept up. We wanted a way to flag confusion as it happens, before frustration turns into giving up, while preserving academic integrity. Learn.AI combines neurotechnology with an “AI mentor” that guides students forward without giving them the answers.

What it does

Learn.AI detects signals associated with cognitive overload/confusion from an EMOTIV headset. When a threshold is reached, it:

Captures a screenshot of the student’s screen.
Sends it to ChatGPT with strict “mentor-mode” guardrails to promote the best educational practices
Returns guidance: clarifying questions, hints, and step-by-step scaffolding—not solutions.

All while built-in policies prevent direct answers and steer the model toward coaching, metacognition prompts, and source-checking.

How we built it

BCI pipeline: EMOTIV streams a signal into our program, accompanied by an action type and confidence level of that action. Given a strong enough confidence and appropriate action, we advance to the LLM.
Mentor mode: OpenAI API with system prompts and rules that enforce hint-first, no-cheating behavior + refusal patterns for answer requests.
Context capture: Secure screenshot capture → optional redaction → vision-enabled prompt for situational guidance.
Stack: Working frontend (React) and backend (Python). EEG labels and confidence → mentor orchestration, and simple admin tools to tune thresholds and guardrails.

Challenges we ran into

EmotivBCI connectivity: Initial device/API quirks blocked real hardware testing. We unblocked development by adapting to virtual brain devices as a strong proof of concept.
Device-specific errors made in team collaboration and the helpful guidance of mentors were a necessity.
Guardrails vs. usefulness: Making the LLM helpful without leaking answers required multiple prompt iterations, role separation, and systematic refusals.

Accomplishments that we're proud of

A full EEG → trigger → screenshot → mentor guidance loop working end-to-end.
Integrity-first design: Guardrails that keep the assistant in true mentor mode.
A demo that feels like a real classroom tool, not just a lab prototype.
Seamless design implementation for users.

What we learned

BCI: signal quality, thresholds, and contact checks matter as much as models.
Prompt-engineering for LLMs: how to structure hints, Socratic questioning, and step sizing.
Production basics: auth flows, error handling, and UX that keeps students in control (opt-in, pause/resume, privacy).

What's next for Learn.AI

Smarter detection: Personalization and adaptive thresholds for different learners/tasks.
Better mentoring: Expand and fine tune the skills of our AI tutors to account for even the hardest of questions.
BCI fidelity: Improve artifact handling and calibration for more accurate confusion triggers.