🚀 AI Lead Scoring Engine — Cleardeals Assignment

An end-to-end AI-powered real-time Lead Scoring System designed to help brokers prioritize high-intent leads by predicting an intent score using machine learning and semantic reranking.

---

📌 Problem Statement

Brokers spend too much time on low-intent leads. Goal: Predict high-intent leads and improve conversion efficiency.

This solution delivers real-time intent scores via API using:

• A trained Gradient Boosted Tree (XGBoost) model
• Optional LLM re-ranking using MiniLM embeddings
• Redis for fast caching
• FastAPI to expose a scoring endpoint
• Async simulation of CRM push integration

---

🧰 Tech Stack

Component	Technology
ML Model	XGBoost
API Server	FastAPI + Uvicorn
Re-Ranker (LLM)	SentenceTransformers (MiniLM)
Cache	Redis
Async Requests	httpx
Dataset	Synthetic (1000 leads)

---

⚙️ Features

• /score → Accepts lead info, returns predicted intent score in <300ms
• Redis caching for repeated leads
• /rerank → Reranks top leads using semantic relevance
• Async CRM push simulation
• Handles edge cases and missing fields gracefully

---

📁 Project Structure

cdt/
├── app.py                 # FastAPI server
├── model_utils.py         # Model loading & prediction
├── llm_reranker.py        # LLM-based lead reranking
├── train_model.py         # Training script (XGBoost)
├── generate_fake_leads.py # Creates 1000-lead dataset
├── leads.csv              # Simulated dataset
├── lead_score_model.pkl   # Trained XGBoost model
├── feature_columns.json   # Feature order mapping
├── requirements.txt
├── .env                   # Optional Redis config
└── README.md

---

🚀 Running the Project

1. Install dependencies

pip install -r requirements.txt

2. Start Redis

redis-server

3. Run the API

uvicorn app:app --reload

4. Access API Docs

Visit: http://127.0.0.1:8000/docs

---

🧪 Endpoints

/score — POST

Predicts intent score in real-time. Accepts lead data.

/rerank — POST

Accepts list of top leads + a target_intent string. Returns re-ranked leads by semantic similarity using MiniLM.

---

🧠 How It Works

• XGBoost model scores leads from tabular features
• Optional LLM (MiniLM) computes semantic similarity for top leads
• Redis caches past results
• Async CRM webhook simulated with httpx

---

🔐 Compliance

• No PII is stored or exposed
• All sensitive fields are excluded from scoring pipeline
• Designed to be DPDP-ready

---

✍️ Author

Dharmik Sompura

📧 Email: [your_dharmiksompura1212@gmail.com]

🔗 GitHub: [https://github.com/Dharmik0712]

🔗 LinkedIn: [https://www.linkedin.com/in/dharmik-sompura/]