Inspiration
To dramatically reduce time spent by focusing on the right skills while avoiding common pitfalls—ensuring we don’t overlook essentials or invest effort in low-impact learning.
What it does
Although this project is in initial stage, idea is that One Gemini Agent Monitors the overall state and which week a question belongs to and changes to the most relevant week or adds new module for classification and judging based on topics
How we built it
This application is powered by a Multi-Agent Orchestration architecture. It has a total of 3 distinct AI Agents working in coordination:
- The Autonomous Learning Agent (The "Teacher") Role: Found in geminiService.ts. It acts as the primary instructor and pedagogical strategist. Intelligence: It uses a "Reasoning Loop" (Planner, Teacher, Assessor, Diagnostician, Verifier, Memory Manager) to decide how to interact with you based on your current mastery levels. Model: Primarily gemini-3-flash-preview with a fallback to gemini-2.5-flash-lite.
- The Curriculum Navigator (The "Router") Role: Found in routingService.ts. It acts as an "Intent Analyzer" that sits between your messages and the UI. Intelligence: It listens to your requests to decide if the app should navigate to a different week or dynamically generate a new module if you ask for a topic outside the 11-week scope (e.g., asking about Quantum Mechanics in a Calculus course).
- The Live Vision Agent (The "Tutor") Role: Found in LiveSession.tsx. It handles the real-time multimodal stream. Intelligence: It "sees" your hand gestures (via the MediaPipe integration) and "hears" your voice commands simultaneously to provide immediate feedback on your air-drawn equations. Model: gemini-2.5-flash-native-audio-preview-12-2025.
Challenges we ran into
Cloud deployment from AI builder could provide even more info in cloud console logs
Accomplishments that we're proud of
Is it "Agentic AI"?
And Yes, absolutely. It is not a standard chatbot; it is a Closed-Loop Agentic System for several reasons: State Persistence (Memory): It maintains a "Neural State" (LearnerState) that it reads from and writes to after every turn. It "knows" what you've mastered and where you've failed.
Autonomous Decision-Making: The model does not just answer questions; it chooses an Action (TEACH, ASSESS, RETEACH, etc.). It self-corrects—if it detects you are confused, it autonomously decides to halt the lesson and "Reteach."
Environment Interaction: The Routing Agent can actually change the environment by adding new weeks to the curriculum or moving the user to different parts of the app, demonstrating agency over the software interface itself.
Verification Step: In every response, it is forced to perform a "Verification" step (Section 4 of its output), where it audits its own mathematical logic before presenting it to you.
What we learned
Gemini's all rounder capability with multimodality and how it could work together with its clones for agentic system.
What's next for Math University
Currently we have just shown in the last part of the demo that there could be a possibility where the index finger would be tracked via mediapipe and we could superimpose the trajectory into a canvas whereby the use could draw as in like a canvas while the user moves the fingers through air. Another Gemini model would be able to differentiate when to add breaks to the superimposed trajectory based on fast and slow movements in drawing and depth of the field of view as generally we take back our hands to draw more.
Also currently there is connect to mobile option with streaming, so that there is extra maneuverability and convenience. It is in partial development and could be developed further.
Log in or sign up for Devpost to join the conversation.