🎓 Campus Placement Intelligence Platform

An AI-powered analytics platform that provides accurate, explainable insights from multi-year college placement data using Retrieval-Augmented Generation (RAG).

---

❓ Problem Statement

Colleges publish placement data across multiple PDFs and CSV files, making it difficult for students to:

• 📊 Find accurate placement statistics
• 📈 Analyze trends across years
• 💬 Get answers to natural language questions

Traditional dashboards:

• Require predefined queries
• Lack flexibility
• Fail to explain results clearly

---

💡 Solution Overview

This project builds an AI-driven placement intelligence system that:

• 📥 Ingests multi-year placement data into a centralized database
• 🧮 Pre-aggregates analytical summaries
• 🤖 Uses a RAG-based chatbot to answer placement-related questions accurately
• 🔍 Ensures transparency via citations and coverage indicators

🚫 The system avoids hallucination by restricting the LLM to verified, indexed data only.

---

🏗️ System Architecture

CSV Files (Year-wise)
        ↓
MySQL (placements table)
        ↓
SQL Aggregation (Summaries)
        ↓
FAISS Vector Store
        ↓
RAG Chatbot (Local LLM)

🔑 Key Design Principles

1. Single normalized database table
2. Pre-aggregated, LLM-friendly summaries
3. Pure RAG (no runtime SQL or tool calling)
4. Admin-controlled refresh pipeline

---

🧰 Tech Stack

Backend

• FastAPI
• SQLAlchemy
• MySQL

AI / NLP

• FAISS (Vector Store)
• Ollama (Local LLM – LLaMA 3)
• LangChain

Data Processing

• Pandas

Other

• Python
• REST APIs

---

🗄️ Database Design

placements

----------------------------------
id  
student_name  
branch  
company  
package_lpa  
academic_year  

Why a Single Table?

• ✅ Enables easy multi-year analysis
• ✅ Simplifies trend computation
• ✅ Scales without schema changes

---

🔄 Data Ingestion Pipeline

1. Year-wise CSV files are placed in data/raw/
2. A loader script automatically ingests all CSVs into MySQL
3. New academic years can be added without code changes

📄 Script:

python scripts/load_to_mysql.py

---

🧠 RAG Design

Instead of querying the database at runtime, the system:

1. Pre-computes analytical summaries (per year + overall)
2. Converts them into natural-language text
3. Stores them in a FAISS vector database

The chatbot answers queries strictly from indexed knowledge.

This ensures:

• ✔️ Correctness
• ✔️ Consistency
• ✔️ Zero hallucination

---

🤖 Chatbot Features

• 💬 Natural language Q&A on placement data
• 📆 Multi-year analytics support
• 🛡️ Placement-only guardrails
• 🔗 Transparent citations
• 📊 Coverage indicator showing data confidence

---

🔐 Admin Refresh Endpoint

An admin-only endpoint allows rebuilding the knowledge base without restarting the server.

POST /admin/refresh

This endpoint:

1. Re-generates summaries
2. Rebuilds FAISS index
3. Updates the chatbot in memory

---

⚙️ Challenges Faced & Solutions

Challenge	Solution
Partial answers from RAG	Introduced pre-aggregated summaries
Incorrect analytics	Removed runtime SQL & tool calling
Year-wise data mismatch	Dynamic summary generation from DB
LLM hallucination	Strict RAG-only answers + guardrails

---

▶️ How to Run Locally

# install dependencies
pip install -r requirements.txt

# load CSVs into MySQL
python scripts/load_to_mysql.py

# start backend
uvicorn app.main:app --reload

---

🌟 Why This Project Stands Out

1. Uses industry-standard RAG architecture
2. Avoids LLM hallucination by design
3. Scales to multiple years without schema changes
4. Demonstrates system design, not just coding

---

👩‍💻 Author

Rucha Chavan
Java Backend | DSA | AI Systems | RAG-based Applications