🎓 Saarthi AI

Intelligent Scholarship Discovery with Live Multi-Agent Reasoning

The first scholarship platform where you can watch AI agents think in real-time

🚀 Live API · 📄 API Docs · 💻 GitHub

---

Built for Convolve 4.0 Hackathon - MERGE-CONFLICT | IIT Roorkee

---

🎯 The Problem

85% of eligible Indian students never find scholarships meant for them.

Challenge	Impact
5,000+ scattered schemes	Information overload
Keyword-only search	Misses semantic intent
No personalization	Same results for everyone
Zero transparency	"Why did I match?" unanswered
Scam-prone ecosystem	Trust deficit

Students search: "गरीब छात्रों के लिए पैसा" (money for poor students) Portals expect: "AICTE Pragati Scheme SC Category 2026"

---

💡 Our Solution

Saarthi AI (Sanskrit: "Guide/Charioteer") transforms scholarship discovery through a transparent multi-agent system where you can watch AI agents reason in real-time.

✨ What Makes Us Unique

🤖 Live Agent Visibility Watch 5 specialized AI agents collaborate on your search. No black box - full transparency. 🔍 Hybrid Search (88% Precision) BM25 + Vector + RRF Fusion combines keyword precision with semantic understanding.

🧠 Memory-Based Personalization Your interactions evolve results (+16% NDCG) with filter bubble prevention. ✅ Zero Hallucinations Deterministic eligibility checking - rule-based, not LLM. Legal criteria never guessed.

---

🏗️ System Architecture

graph TB
    subgraph Frontend["🖥️ Frontend (Next.js 14)"]
        UI[React UI] --> WS[WebSocket Client]
        UI --> Search[Search Interface]
    end

    subgraph Backend["⚙️ Backend (FastAPI)"]
        API[REST API] --> Orchestrator[Agent Orchestrator]
        WSS[WebSocket Server] --> Orchestrator

        subgraph Agents["🤖 Multi-Agent System"]
            QA[Query Agent] --> RA[Retrieval Agent]
            RA --> MA[Memory Agent]
            MA --> EA[Eligibility Agent]
            EA --> ResA[Research Agent]
        end

        Orchestrator --> Agents
    end

    subgraph Storage["💾 Qdrant Vector DB"]
        Scholarships[(200 Verified Scholarships)]
        Memory[(User Interactions)]
    end

    Frontend <--> Backend
    Backend <--> Storage

Loading

Agent Pipeline Flow

Agent	Purpose	Latency
🔍 Query Agent	Parse and understand search intent	5ms
📚 Retrieval Agent	Hybrid BM25 + Vector search with RRF fusion	25ms
🧠 Memory Agent	Personalize based on user history	8ms
✅ Eligibility Agent	Deterministic 100-point scoring	15ms
🌐 Research Agent	Fallback web search (18% of queries)	45ms

---

📊 Performance Metrics

Metric	Value	Significance
P95 Latency	59ms	Real-time UX
Precision@5	88%	High relevance
Memory Impact	+16% NDCG	Proven personalization
Eligibility	100% Deterministic	Zero hallucinations
Dataset	200 verified	From .gov.in sources

---

🛠️ Tech Stack

Layer	Technology
Frontend	Next.js 14, React 18, TypeScript, TailwindCSS, Framer Motion
Backend	Python 3.10+, FastAPI, Pydantic, uvicorn
Vector DB	Qdrant (hybrid search, dual collections)
Embeddings	sentence-transformers (all-MiniLM-L6-v2)
Search	BM25 + Vector + RRF Fusion
Real-time	WebSocket (live agent streaming)

---

🚀 Quick Start

Prerequisites

• Python 3.10+
• Node.js 18+
• Docker (for Qdrant)

1. Clone & Setup

git clone https://github.com/krishdef7/Saarthi-AI.git
cd Saarthi-AI

2. Start Qdrant

docker run -d -p 6333:6333 -v $(pwd)/qdrant_storage:/qdrant/storage qdrant/qdrant

3. Backend Setup

cd mas_scholar_app/backend

# Create virtual environment
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Copy environment file
cp ../../.env.example .env

# Start API server
uvicorn main:app --reload --port 8000

4. Frontend Setup

cd mas_scholar_app/frontend

# Install dependencies
npm install

# Set environment variable
export NEXT_PUBLIC_API_URL=http://localhost:8000

# Start development server
npm run dev

5. Open in Browser

Navigate to http://localhost:3000 and try:

• "engineering scholarship for SC category"
• "merit based postgraduate scholarship"
• "PMSS-2026" (exact scheme ID)

---

📁 Project Structure

saarthi-ai/
├── README.md                          # This file
├── SUBMISSION.md                      # Hackathon submission
├── Saarthi_AI_Final_Submission.pdf    # Technical report
├── docs/
│   ├── ARCHITECTURE.md                # System design
│   └── METRICS.md                     # Performance benchmarks
│
├── mas_scholar_app/
│   ├── backend/                       # FastAPI server
│   │   ├── main.py                    # API entry point
│   │   ├── requirements.txt           # Python dependencies
│   │   ├── data/                      # 200 verified scholarships
│   │   ├── models/                    # Pydantic schemas
│   │   ├── routers/                   # API endpoints
│   │   └── services/                  # Core business logic
│   │       ├── hybrid_search.py       # BM25 + Vector + RRF
│   │       ├── eligibility.py         # Deterministic scoring
│   │       ├── user_memory.py         # Vector memory
│   │       └── websocket.py           # Live streaming
│   │
│   └── frontend/                      # Next.js 14 app
│       ├── src/
│       │   ├── app/                   # App router pages
│       │   ├── components/            # React components
│       │   └── hooks/                 # Custom hooks
│       └── package.json
│
├── render.yaml                        # Backend deployment
├── docker-compose.yml                 # Local development
└── .env.example                       # Environment template

---

🎨 Key Features

🤖 Live Agent Pipeline

Watch agents collaborate in real-time via WebSocket. See exactly what each agent does.

🔍 Semantic Search

Natural language queries work. Hindi support. Handles both "scholarship for poor" and "AICTE-2026".

📊 Eligibility Radar

Visual breakdown of how you match each scholarship across 5 dimensions.

🛡️ Trust Scoring

23 pattern-based scam detection. Government sources preferred.

💾 Persistent Memory

Your interactions improve future results (with decay and diversity safeguards).

---

🏆 Hackathon Alignment

Criterion	Our Implementation
Originality	First scholarship system with live agent visibility
Problem Understanding	Deep research into India's scholarship gap
Innovation	Hybrid search + vector memory + deterministic eligibility
Technical Implementation	Production-ready, <60ms latency
GitHub Quality	Clean structure, meaningful commits
Documentation	Comprehensive docs, diagrams, metrics
Design & UX	Modern UI, mobile-responsive, accessible

---

📜 License

MIT License - See LICENSE file

---

🙏 Acknowledgments

• Qdrant for the excellent vector database
• National Scholarship Portal for public data
• Open Source Community for FastAPI, React, sentence-transformers

---

Built with ❤️ for India's students

सारथी AI - आपका शिक्षा मार्गदर्शक

---

⬆ Back to Top