Statusshipped
CategoryAI RAG LLM Multi-Agent Systems Education Technology Vector Search AI Tutor
Updated2026-03-29T16:47:54.166686+00:00
Timeline5 steps
Overview
Overview
MeroStudySathy is an AI-powered learning system that transforms static PDF documents into interactive, guided learning experiences.
Instead of passively reading long documents, users receive a structured study plan, AI-guided explanations, practice questions, and detailed feedback — all generated directly from the uploaded material.
The system uses a multi-agent architecture where specialized AI agents collaborate to analyze documents, teach concepts, generate practice problems, and evaluate answers.
Key Features
• Multi-agent architecture (Planner, Teacher, Practice, Evaluator)
• Retrieval-Augmented Generation (RAG) pipeline
• Vector search using document embeddings
• AI-generated structured study plans
• Interactive teaching sessions with document citations
• Automatic practice question generation
• Answer evaluation with feedback and scoring
• Weak topic detection and progress tracking
• Response caching to reduce repeated API costs
How It Works
- User uploads a PDF document
- The system extracts text and splits it into semantic chunks
- Embeddings are generated and stored in a SQLite vector database
- A Planner Agent analyzes the document and generates a learning roadmap
- Teacher Agent explains concepts using structured lessons
- Practice Agent generates quizzes
- Evaluator Agent scores answers and tracks progress
Architecture
The system follows a local-first Retrieval-Augmented Generation (RAG) architecture where documents are processed, indexed, and taught through specialized AI agents.
┌─────────────────────────────────────────────────────────────────┐
│ YOUR PDF DOCUMENT │
└────────────────────────────┬────────────────────────────────────┘
│
▼
┌────────────────┐
│ PDF EXTRACTION │
│ (per-page) │
└────────┬───────┘
│
▼
┌────────────────┐
│ TEXT CHUNKING │
│ 1000 tok/150 │
└────────┬───────┘
│
▼
┌────────────────┐
│ EMBEDDINGS │
│ (batch: 100) │
└────────┬───────┘
│
▼
┌────────────────────────┐
│ SQLITE VECTOR STORE │ ←── Response Cache Layer
│ (local database) │ (0 API cost on repeat)
└───────────┬────────────┘
│
┌───────────────────┼───────────────────┐
│ │ │
▼ ▼ ▼
┌─────────┐ ┌─────────┐ ┌─────────┐
│ PLANNER │ │ TEACHER │ │PRACTICE │
│ AGENT │ │ AGENT │ │ AGENT │
└────┬────┘ └────┬────┘ └────┬────┘
│ │ │
▼ ▼ ▼
Learning Plan Teaching Sessions Quiz Questions
(structured) (with citations) (with feedback)
│ │ │
└──────────────────┼───────────────────┘
│
▼
┌─────────────┐
│ EVALUATOR │
│ AGENT │
└──────┬──────┘
│
▼
Progress Tracking
Weak Topic Identification
All documents, embeddings, and response caches are stored locally in a SQLite database, ensuring privacy while significantly reducing API costs through response caching.
Why I Built This
Most study tools only highlight or summarize text. I wanted to build something closer to a real tutor — a system that understands a document and teaches it interactively.
This project explores AI agent collaboration, retrieval systems, and educational workflows to make self-learning more effective.
Key Features
- AI RAG LLM Multi-Agent Systems Education Technology Vector Search AI Tutor
- Next.js 14
- TypeScript
Project Timeline
Step 1
Concept & Problem Research
Identified the limitations of traditional PDF-based studying and designed a system to convert static documents into interactive AI tutoring sessions.
Step 2
RAG Pipeline Implementation
Built the document processing pipeline including PDF extraction, semantic chunking, embeddings generation, and vector search using SQLite.
Step 3
Multi-Agent Learning System
Implemented Planner, Teacher, Practice, and Evaluator agents to simulate a real tutoring workflow.
Step 4
Interactive Learning Interface
Developed a real-time teaching interface with streaming responses, source citations, and follow-up questioning.
Step 5
Evaluation & Progress Tracking
Added automated answer evaluation, scoring, and weak topic detection to track learning progress.