An advanced, secure, and personalized educational AI tutor designed with a multi-agent hierarchy using the Google Agents SDK.
Traditional learning systems are rigid, treating class cohorts as uniform blocks. When students study, they often focus on what they are already good at, leaving critical blind spots unaddressed.
The Trifecta Agent solves this problem dynamically. By integrating a turn-by-turn interactive diagnostic quiz, analyzing performance details programmatically, and combining indexed study material via a custom FastMCP server, the system crafts personalized, high-quality 7-day study paths.
To bridge safety and education, we designed a multi-step moderation and security pipeline. Presidio-driven PII scrubbing ensures data privacy, while dual Human-in-the-Loop approval layers ensure that parents, teachers, and moderators remain fully in control of the educational journey.
Dynamically adjusts question difficulty based on rolling score accuracy, preventing discouragement while ensuring a rigorous test.
Grounds generated content using a FastMCP-powered retrieval system. Ingests PDFs and URLs into a local TF-IDF semantic index.
Scrub PII instantly using Microsoft Presidio and preemptively block prompt-injection attacks with integrated guardrails.
Generates deterministic 7-day schedules focusing explicitly on flagged weakness topics, complete with offline PDF exports.
Pauses workflow execution to request parent, teacher, or moderator approval before resolving critical educational decisions.
Provides dual UIs: A responsive React SPA for students and a comprehensive FastAPI glassmorphic console for teachers, plus seamless CLI testing via the ADK Playground.
Lead AI Architect
Spearheaded the design of the multi-agent graph hierarchy, state models, and structured orchestrations using the Google Agents SDK.
Frontend Systems
Crafted the responsive SPA student workspace portal, integrated the FastAPI backend layers, and polished the responsive telemetry layouts.
Security Engineer
Engineered the Presidio protection safeguards, integrated FastMCP vector store ingestion, and designed secure human review interrupts.
This application was engineered as a capstone submission for the Kaggle course. We are extremely grateful for the support and guidance of our mentors and communities:
For designing comprehensive training tracks, providing computational test beds, and hosting this capstone framework challenge.
For building the robust Google Agents SDK framework and the ADK tooling system that powered our orchestration graph.
For contributing the packages that enabled Presidio analyzer scrubbers, Vite, React, ChartJS, and beautiful styling tools.