I am open to new roles.Dallas–Fort Worth, TX

Kamal Mehta

Machine Learning Engineer

I built machine learning that runs in production on 500+ locomotives.

Built machine learning for 5+ years, 4 of them at Progress Rail (Caterpillar). Shipped autonomous-train ML for a Class I U.S. freight railroad and an Australian client.

Fuel saving vs manual driving on trips from Aug 2024 to Dec 2025; the software is EPA-certified (2025).: 12%
Per model: a 2 GB tree ensemble → one 15–20 MB file, on 500+ locomotives.: ~100×
Fleet availability, Jan 2024 to mid-2026, while I closed 100+ field issues, each with a regression test.: ~90% → 99%

Resume Contact

Selected work

Projects

Twenty-one projects, strongest first for machine learning work. Open a card for the problem, how it works, and the numbers.

More work

This page draws its project cards with JavaScript. With scripts off, the same projects are listed here as text.

Deep RL Driving Strategy for Freight Trains

Progress Rail · 2023 – 2026

Shipped DQN driving, with a slack predictor for jolts between cars, to 500+ locomotives.
100+ Field Issues Investigated, Each With a Regression Test

Progress Rail · 2024 – 2026

Built a Lambda reliability pipeline, a fleet-health crawler, and a regression test per failure.
One Model per Train Type, Not per Territory

Progress Rail · 2024 – 2025

Engineered look-ahead terrain features so one RL model serves a train type anywhere.
Making the Simulator Agree With the Real Train

Progress Rail · 2022 – 2023

Fit the simulator to field trips with an exhaustive bounded grid search.
Floor Plans from Handheld Scans: About 95–98% Wall-Position Accuracy

Artifex Technologies, Inc. · 2021 – 2022

Point clouds plus IMU; density-based clustering finds walls as the user walks.
Object Detection & Camera–LiDAR Fusion on an Unmanned Vehicle

Graduate Research · Computer Vision & Sensor Fusion · 2020 – 2021

YOLOv3 trained on a hand-built dataset, fused with LiDAR through a Kalman filter.
Double DQN Eco-Driving With a Slack Filter, in PyTorch and TensorFlow

Personal project · public code · 2026

A Double DQN drives a simulated train; a slack predictor vetoes risky actions.
The Data Pipeline Behind an Australian Railroad Client's Rollout

Progress Rail · 2024

Fused two recorders into one per-second record with GIS track profiles; large gaps split trips.
Sarathi: A Private, Local-First AI Board With Agents and a Voice Assistant

Personal project · 2026

Local LLM agents and Mitra, a three-language voice assistant, scored on fresh held-out sets.
Disha: A Local-First Career Profile Reviewer That Scores Against Live Job Postings

Personal project · 2026

Scores a resume, LinkedIn, GitHub and site for one role, against live postings.
Building Train Make-Ups (Consists) in Parallel

Progress Rail · 2024

Parallelized train consist builds with Python multiprocessing over OpenSearch.

The full record is in the resume (PDF).

The short version

Fifteen stops in five minutes

9 of the 10 projects, plus career, education, certificates and beyond work. Silent, drawn in code. Pause, skip or go back at any point.

Take the 4-minute tour

How I work

Find the real problem first, then build

Career path

Experience

Toolbox

Skills

Every skill links to the projects that used it.

Always learning

Certificates

The graduate certificate, then courses. Verify opens the issuer's page; View certificate opens the scan.

Background

Education and thesis

Off the clock

Beyond work

What the rest of the day looks like.

What I believe

Let's talk

Write to me about a role or the work

Open to Machine Learning Engineer and Perception Engineer roles. Reach out any way that is easiest for you.