Inspiration
There are many other dimensions to learning besides simply answering questions correctly. Confidence vs uncertainty is a major one, but there are other elements of how a student answers a question that meaningfully inform the tutor's approach. This dynamic is especially apparent when learning via the Feynman Technique, where a student essentially tries to progressively teach a concept to a 12-year-old.
How can the tutor of tomorrow use technology to get more signal into their students' progress through a lesson?
What it does
An AI tutoring agent prompts the student with a question about a STEM topic. The student types their response in text (requiring concision and clarity of understanding) while speaking through their thinking out loud (evaluating their confidence and natural chain of thinking).
The correctness of the student's answer is evaluated by a reasoning LLM looking at the text of the answer. The confidence of the student's answer is measured by an ELM (Ensemble Listening Model) that considers up to 26 elements of the student's voice. Features like language, prosody, timing, and emotion all contribute to a score that the agent can use to adjust its tutoring pattern.
The supervising human instructor can consider these 2 dimensions of learning when reviewing a student's progress. The more interaction with the tutoring agent, the more aggregate confidence there is in the student's level of understanding.
How we built it
FastAPI (Python) backend Next.js frontend SQLite (aiosqlite) DB Modulate Velma (voice analysis) Google Gemini API -- gemini-2.5-flash (reasoning/inference) Lightdash (analytics)
Challenges we ran into
Latency, simplicity of UX
Accomplishments that we're proud of
Novelty?
What we learned
Having a clear vision for the medium and integration with tools helps shape process.
What's next for Tutor
Deploy, expand knowledge areas, add support for instructor-level input
Log in or sign up for Devpost to join the conversation.