Skip to main content
Anugrah E.

EduGuideX

Follow one student's journey to understand how personalized, local AI tutoring actually works.

Scroll to Begin

Alex

High School Sophomore

Struggles
  • Physics concepts
  • Needs personalized pacing

Meet the Student

Education isn't one-size-fits-all. Meet Alex.

Alex has goals, but struggles with rigid curriculums. When Alex hits a mental block, standard platforms either offer zero help or simply provide the final answer, bypassing the actual learning process.

Setting a Goal

Alex types in a simple goal.

The interface accepts the request. There's no complex dashboard, no intimidating AI jargon. Just a soft glow indicating that the system is processing the goal and preparing a personalized journey.

Learning Goal

Finding the Right Mentor

Instead of introducing a generic AI, the system selects specialized sub-agents.

One central orchestrator splits into distinct personalities: The Socratic guide to ask questions, the Knowledge builder to provide facts, and the Forecaster to predict mastery. Each appears naturally, only when needed.

Socratic Agent
Guides without answering
Knowledge Agent
Provides strict factual context
Forecast Agent
Predicts learning outcomes
Course Agent
Maps curriculum milestones
Socratic Guide
Asking questions, not giving answers
If velocity is the rate of change of position, what happens when velocity is constant? Does position change?
Yes, the position still changes, but it changes at a steady rate.
Exactly! And if the rate of change is steady, what does that mean for acceleration?
Concept Mastery: KinematicsGrowing

Understanding the Student

True learning happens in the struggle.

The AI asks probing questions instead of spoon-feeding facts. As the student responds, the system's internal knowledge graph updates dynamically. The visitor experiences personalized learning rather than just reading about it.

Building a Learning Plan

Based on the interaction, the system generates a structured roadmap.

The student receives weekly milestones, recommended resources, and SMART goals tailored exactly to their pace. The system adapts as the student's progress changes.

Week 1
Mastering Kinematics
Focus on velocity and acceleration formulas.
Week 2
Newton's Laws
Understanding forces in equilibrium.
Week 3
Practice Exam
Applying concepts to multi-step problems.
Progress Dashboard
Live Sync
Practice Exam Score
85%
+25% increase
Topic Mastery
Kinematics
Mastered
Newton's Laws
Improving
Forecast ConfidenceHigh

Tracking Progress

Learning is invisible until it's measured.

As Alex continues to interact with the Socratic guide and complete milestones, assessment scores improve. Weak topics visually transition into mastered subjects, and the system's forecast confidence rises. The product delivers on its promise.

Behind the Intelligence

The student experiences a seamless journey. But under the hood, how does a raw student request turn into a personalized Socratic question? We peek behind the curtain.

Student Request
"I don't understand velocity."
Express Router & Controller
Routes to the specific agent based on intent.
Ollama + Local LLM (Qwen)
Processes prompt locally for privacy and speed.
Structured JSON Output
{ "type": "socratic_question", "content": "..." }
Application Response
Renders as a seamless chat bubble in the UI.

System Overview

The architecture is designed to support the journey seamlessly. See how the underlying technology maps to the product experience.

Student
Goal
Learning
Assessment
Roadmap

My Contribution

Product Vision & Scope
Intelligent Tutoring Research
Multi-Agent Workflow Design
Ollama + Local LLM Integration
Agent Responsibilities Definition
AI-Assisted Implementation
Prompt Engineering
JSON Validation Handling
Iterative Refinement

Development Methodology

Built strictly during a 24-hour hackathon. Time constraints meant relying heavily on AI coding assistants for frontend boilerplate so I could focus my mental energy on the core differentiator: orchestrating multiple specialized agents and getting Ollama to respond reliably.

Hardest Challenge

Forcing a small, quantized local LLM (Qwen) to output strict, parseable JSON for application state. The model naturally wants to output conversational text. Extensive prompt engineering and fallback validation were required to prevent the app from crashing on malformed responses.

What I Learned

A single "god prompt" fails terribly for complex educational tasks. Decomposing the system into micro-agents (Socratic, Forecast, Knowledge) dramatically improved response quality and predictability.

Future Improvements

Implementing RAG (Retrieval-Augmented Generation) with a vector database for long-term student memory, allowing the AI to reference a struggle the student had three weeks prior.