QuoteFlow AI 🤖

Autonomous Quote Agent Pipeline

A multi-agent AI system that autonomously processes auto insurance quotes end-to-end — profiling risk, predicting conversion, advising on premiums, and routing decisions without human intervention.

---

Problem Statement

Auto insurance carriers generate thousands of quotes daily. Yet only 1 in 5 quotes converts to a bound policy. The rest expire silently — representing millions in wasted acquisition spend. Every unconverted quote currently demands a human to investigate manually.

QuoteFlow AI eliminates that bottleneck.

---

The 4-Agent Pipeline

Agent	Model	Mode	Role
01 — Risk Profiler	XGBoost 2.0	Fully Automatic	Analyses accident history, citations, driving experience, age, vehicle usage → outputs LOW / MEDIUM / HIGH risk tier with SHAP explanation
02 — Conversion Predictor	LightGBM + SMOTE	Fully Automatic	Scores each unbound quote 0–100% bind probability. 3-layer imbalance fix handles the 22% class skew
03 — Premium Advisor	Llama 3.1 8B (Ollama)	Hybrid	Reasons whether quoted premium is a conversion blocker by comparing it to the customer's salary bracket using chain-of-thought LLM reasoning. No API key required
04 — Decision Router	Rules + Llama 3.1	Escalate-Only	Combines all upstream outputs → routes to Auto-Approve (~45%), Follow-Up (~35%), or Escalate (~20%). Only escalations ever touch a human underwriter

---

Tech Stack

LangGraph        → Agent orchestration & stateful pipeline
XGBoost          → Agent 1: Risk profiling
LightGBM + SMOTE → Agent 2: Conversion prediction (22% imbalance handled)
Llama 3.1 8B     → Agent 3 & 4: Local LLM reasoning (via Ollama)
SHAP             → Explainability at every agent node
FastAPI          → Backend REST API
Vanilla JS       → Frontend dashboard
Chart.js         → Live updating charts

---

How to Run

Prerequisites

• Python 3.12+
• Ollama installed
• VS Code with Live Server extension

Step 1 — Create Virtual Environment

python -m venv venv

Step 2 — Activate Virtual Environment

# Windows
venv\Scripts\activate

# Mac/Linux
source venv/bin/activate

Step 3 — Install Dependencies

pip install -r requirements.txt

Step 4 — Start Llama 3.1 (Local LLM)

ollama run llama3.1

Keep this terminal open. First run downloads the model (~4.7GB). Subsequent runs are instant.

Step 5 — Start Backend (New Terminal)

cd backend
uvicorn main:app --reload --port 8000

Backend runs at http://127.0.0.1:8000

Step 6 — Start Frontend

Right-click frontend/index.html → Open with Live Server

Frontend runs at http://127.0.0.1:5500/frontend/index.html

---

Project Structure

QuoteFlow_AI/
├── backend/
│   ├── data/
│   │   └── quotes.csv          ← 146,259 real insurance quotes
│   ├── models/
│   │   ├── xgb_risk.pkl        ← Trained XGBoost risk model
│   │   ├── lgb_conv.pkl        ← Trained LightGBM conversion model
│   │   ├── encoders.pkl        ← Feature encoders
│   │   ├── conv_features.pkl   ← Conversion feature list
│   │   └── risk_features.pkl   ← Risk feature list
│   ├── main.py                 ← FastAPI endpoints
│   ├── pipeline.py             ← LangGraph 4-agent pipeline
│   ├── explainer.py            ← SHAP explainability
│   ├── train_models.py         ← Model training script
│   └── requirements.txt
├── frontend/
│   ├── index.html              ← Main dashboard
│   ├── app.js                  ← Dashboard logic
│   └── style.css               ← Styling
└── README.md

---

Escalation Triggers

A quote escalates when any one condition is met:

- Bind score < 35%
- Risk = HIGH and Bind score < 50%  
- Premium BLOCKER + Re-Quote + HIGH risk
- Model confidence < 60%

---

Dataset

• 146,259 real auto insurance quotes
• 22.2% conversion rate (1 in 5 quotes binds)
• 8 regions (A through H), EA and IA agent types
• 3-layer imbalance fix: SMOTE + class_weight='balanced' + threshold tuning

---

Built For

GITAM AI DAY Hackathon — Use Case 3: Autonomous Quote Agents