Members-Only
Recent Talks & Demos are for members only
You must be an AI Tinkerers active member to view these talks and demos.
Supervised Socratic AI Tutors
Learn how a multi‑agent tutoring system uses Socratic dialogue, supervisory control, and curriculum‑grounded RAG to guide student reasoning and prevent answer leakage, with proven outcomes.
This talk presents a novel multi-agent intelligent tutoring system that prioritizes student reasoning over answer-giving. Unlike typical LLM-powered tutors that default to information delivery, our framework implements a supervised architecture where specialized agents work in concert: a Socratic agent conducts grade-appropriate dialogues, a supervisor ensures pedagogical alignment and prevents answer leakage, an evaluation layer tracks reasoning quality through milestone progression, and a RAG component provides curriculum-grounded hints while maintaining strict knowledge boundaries.
The system enforces a misconception-based learning workflow where students must diagnose, explain, and correct errors using evidence—with the AI learning only from student input. An 8-week classroom study across grades 5-10 (627 students, 10 teachers) demonstrated statistically significant learning gains (p=.03), with teachers reporting substantial improvements in evidence-based reasoning and productive struggle.
This session will cover the technical architecture, the challenge of separating student-taught knowledge from RAG-retrieved context, the orchestration loop that regenerates responses based on supervisory feedback, and practical insights from real classroom deployment. Attendees will learn how to build AI systems that elevate—rather than bypass—student thinking.