M3ATRON (Multi-Agentic AI Tutoring System)
Led backend, database, and deployment work for a team final-year project. The AI tutor streams answers grounded in course materials and revisits prerequisites when a learner gets stuck.
Trade-offCourse retrieval lets instructors update material without retraining the model. Grounding depends on what they upload and what search retrieves.
Our team built a tutoring system with separate agents for teaching, challenging misconceptions, and revisiting prerequisites. I led the backend, database, and deployment work, connecting the agents to course retrieval and a streaming interface.
Problem Statement
Generic AI tutors lack domain awareness, hallucinate course materials, cannot adapt pedagogically, and fail to challenge student misconceptions in academic settings.
Engineering Solution & Architecture
Multi-Agent Backend Orchestration: Architected a 3-tier asynchronous system using FastAPI, LangGraph, and Supabase (pgvector) to coordinate specialized agents (Teacher, Devil's Advocate, Backtrace).
Course-Grounded RAG Pipeline: The document service indexes course PDFs and returns relevant passages with their source and page. The tutoring service supplies that context to the model.
Intelligent Intent Routing & Backtracing: Designed an intent classification router that tracks student confusion count (threshold ≥ 2) and automatically rewrites queries to teach foundational prerequisites.
Real-Time Token Streaming: Streamed responses using Server-Sent Events, with separate events for the responding persona, answer text, and citations.