EigenCoach | Devpost

Inspiration

Joe, one of our members, used to teach in public schools in Brazil, where he grew up. He witnessed first-hand the biggest struggle students there face: exam prep for the national examination.

Every year, over 7 million students prepare for the high-stakes Brazilian ENEM exam each year. But public school students face a massive resource gap: the lack of access to the expensive private tutors and quality prep materials.

But it's not just about ENEM.

The truth is, big platforms like Khan Academy or Udemy don't have classes for smaller exams from more marginalized corners of the world. So what if we can generate an AI tutor for ANY exam in the world?

What it does

Eigen Coach is a personal, AI-powered tutor that manages a student's entire exam prep journey.

It Plans: A student signs up, and the Initializer agent creates a personalized, time-aware study schedule based on their exam date. The agent learns about the exam and update knowledge base using the material student submits.

It Quizzes: The Questioner agent selects a targeted question from our MySQL database based on the day's topics and the student's current skill level.

It Tutors: The Chatter agent engages the student in a Socratic conversation. It's designed to never give the direct answer, but instead guides the student to discover the solution themselves, just like a real tutor would.

It Tracks: The Finalizer agent analyzes the entire conversation, scores the student's performance on a 0-100 scale, and automatically updates their skill profile in the database. This creates a closed loop where the tutor gets smarter about the student's needs after every single session.

How we built it

We built Eigen Coach on a robust, asynchronous Python backend using FastAPI.

The core of our project is a Clean 4-Agent System that manages the entire student lifecycle without a complex central orchestrator. We use Claude-Sdk as the foundation of our multi-agent model.

All data—from the question bank to student memory and skill levels—is stored in a unified MySQL database.

The agents interact with the database safely and efficiently through a single MCP server that exposes all necessary tools (e.g., get_question_by_topic, update_skill_level).

We used Pydantic to ensure all our models are type-safe and Docker to containerize our MySQL instance for easy setup and development.

Challenges we ran into

Prompting the Socratic Tutor: Our biggest challenge was engineering the Chatter agent's prompts. It's incredibly difficult to make an LLM guide a student to an answer (the Socratic method) instead of just providing it. This required dozens of iterations to ensure it was helpful without becoming a cheat-sheet.

Stateful Conversations: Managing the chat state for the Chatter agent was complex. We had to ensure each conversation was cleanly scoped to a single question and that the agent had the correct session context (the question, the answer, and the session_id) for every turn.

Agent Coordination: Designing the 4-agent system to work "cleanly" without a central orchestrator was tough. We had to be very disciplined about giving each agent a single responsibility and using the MySQL database as the central source of truth.

Accomplishments that we're proud of

The Socratic Chatter Agent: We are incredibly proud of the Socratic tutoring. It's the core of our "why"—it doesn't just give answers, it truly teaches. Seeing it successfully guide a user from "I don't know" to the correct answer is the magic of this project.

The "Closed-Loop" Learning System: The most powerful feature is our full-circle adaptive loop. The Finalizer agent evaluates a student's chat, and its output (a new skill score) is immediately used by the Questioner agent to select the next question. The system adapts in real-time to the student's growth.

The Unified Database Architecture: Committing to a single MySQL database for everything (questions, student data, memory, skills) was a great decision. It makes the entire system robust, scalable, and easy to query.

What we learned

The Nuance of AI in Education: We learned that the true power of AI in education isn't just information retrieval, but guided discovery. Building a Socratic tutor is far more challenging, and ultimately more valuable, than building a simple Q&A bot.

The Database as the "Brain": We learned to use a unified database as the central "brain" and state-holder for a multi-agent system. It's far more robust and scalable than passing complex JSON objects between API calls.

Agent-Based Design: We gained deep experience in designing "clean" agents with single responsibilities. This separation of concerns (e.g., Questioner only picks questions, Finalizer only grades) makes the whole system more resilient.

What's next for EigenCoach

Implement a Reinforcement Learning Recommender: We will evolve the Questioner agent from a rule-based selector into a true AI strategist. By implementing a Deep Q-Network (DQN), the agent will learn the optimal policy for which question to ask next to maximize a student's long-term learning.