Fleet Risk ML — IIT Bombay TECHFEST 2022 Winner 🏆

A machine learning system for privacy-preserving fleet risk scoring built using TensorFlow Federated. Predicts driver risk, route risk, and vehicle risk from distributed telematics data without centralizing sensitive trip information.

🥇 1st Place — IIT Bombay TECHFEST 2022 Fleet Risk ML Contest Competed against 100 global teams. Led a 3-person engineering team.

---

Problem

Fleet operators want to identify high-risk drivers, routes, and vehicles to reduce accidents and insurance costs. But the telematics data needed for risk scoring — driving patterns, locations, behaviors — is highly sensitive. Centralizing it creates privacy and regulatory risk.

Solution: Federated learning. Train models locally on each vehicle's data, aggregate only model updates (not raw data) into a global risk-scoring system.

---

Three Risk Models

The system trains three complementary models, each evaluating a different dimension of fleet risk:

Model	File	Predicts
Driver Risk	drivermodel.py	Risk score based on driving behavior (acceleration, braking, speed patterns)
Route Risk	routemodel.py	Risk score based on route characteristics (traffic, road quality, weather exposure)
Vehicle Risk	vehiclemodel.py	Risk score based on vehicle telemetry (engine health, maintenance signals)

Each model is trained via TensorFlow Federated, so raw telematics data never leaves the source vehicle.

---

Why Federated Learning?

Traditional ML for fleet risk requires centralizing data from every vehicle into one server — creating:

• Privacy risk — driving patterns can identify individuals
• Regulatory exposure — GDPR, regional data sovereignty laws
• Bandwidth cost — telematics data is high-volume
• Single point of failure — central server breach exposes all fleet data

Federated learning solves all four. Each vehicle (client) trains locally; only encrypted model weights are aggregated. The global model improves without ever seeing raw trips.

---

Architecture

---

Tech Stack

Component	Technology
ML Framework	TensorFlow + TensorFlow Federated
Language	Python 3.8+
Visualization	Tableau
Data	Simulated telematics dataset (contest provided)

---

Repository Structure

techfeast-fleet-risk-ml/

• DSP1/ # Data simulation / preprocessing pipeline 1
• DSP2/ # Data simulation / preprocessing pipeline 2
• drivermodel.py # Federated driver-risk model
• routemodel.py # Federated route-risk model
• vehiclemodel.py # Federated vehicle-risk model
• README.md

---

Getting Started

Prerequisites

• Python 3.8+
• TensorFlow 2.x
• TensorFlow Federated

Install

pip install tensorflow tensorflow-federated pandas numpy scikit-learn

Run a model

python drivermodel.py
python routemodel.py
python vehiclemodel.py

---

Results

• 1st place out of 100 global teams at IIT Bombay TECHFEST 2022
• Successfully demonstrated federated training across simulated multi-vehicle clients
• Generated Tableau dashboard surfacing high-risk routes and driver patterns to fleet operators
• Preserved data privacy throughout — no raw telematics centralized

---

Team

Led a 3-person engineering team through:

• Federated learning architecture design
• Model training and tuning across three risk dimensions
• Dashboard design and final presentation

---

License

MIT — see LICENSE