About Jeremy Castagno

Jeremy Castagno: senior engineer working on computer vision, safety-critical autonomy, and production ML. Work history, projects, and awards.

Summary

I take cyber-physical and machine learning systems from prototype to production, mostly in computer vision and safety-critical autonomy. Today I’m a Senior Engineer at Ride with GPS, building ML and geospatial systems for route planning and navigation. I like owning a problem end to end: working out which variables actually matter, pulling signal from many different sources, and picking the simplest design I can trust.

Before that, I did a PhD in robotics at the University of Michigan on urgent landing for drones, built in-flight LiDAR perception at NASA Langley, spent a summer on visual positioning for autonomous aircraft landing at Reliable Robotics, and taught computer science as an assistant professor. I started my career on refinery safety systems at Valero Energy. Full CV or Resume.

Education

PhD, Robotics, University of Michigan (2021)
MSc, Robotics, University of Michigan (2018)
BSc, Chemical Engineering with minors in Computer Science and Mathematics, Brigham Young University (2013)

Work Highlights

Senior Engineer, Ride with GPS 2023–Present
  • Built an ML model that classifies routes by activity (road, gravel, mountain bike, hiking, running) with 95% accuracy
  • Designed spatial matching between GPS trips and routes, which doubled completed route reviews
  • Built a global road-surface model that improved paved/unpaved accuracy by more than 10%
  • Built pipelines that refresh global snow, wind, and air-quality map layers
Assistant Professor, Springfield College 2021–2023
  • Taught machine learning, C++, data structures, databases, and an intro robotics course
  • Rebuilt parts of the CS curriculum around modern practice: Linux, version control, testing, and profiling
Computer Vision Engineer (Summer Contract), Reliable Robotics 2022
  • Developed visual positioning software that localizes an aircraft relative to the runway using computational geometry and classical computer vision
  • Evaluated it on real flight data, showing high accuracy and >99% availability
Robotics PhD Researcher, University of Michigan 2016–2021
  • Presented methods to incorporate rooftops as emergency landing sites for drones
  • Conducted multi-city analysis predicting roof shapes using ML with sensor fusion
  • Created Polylidar3D, an open-source library that extracts polygons from image and LiDAR data about 100x faster than prior CPU methods
  • Processed over 10 TB of imagery and LiDAR using Docker, Slurm, OpenMP/MPI, and PyTorch
  • Presented this work as an invited speaker at Amazon re:MARS (video)
Computer Vision/LiDAR Engineer, NASA Langley 2019–2021
  • Developed real-time perception that uses onboard LiDAR to find emergency landing sites, running on an embedded GPU and proven in flight tests
  • Integrated the perception software with ROS-based autonomy and evaluated it on experimental data
Process Control Engineer, Valero Energy 2013–2015
  • Led design and verification of safety systems for four gas-fired heaters through simulation, hardware-in-the-loop, and field testing
  • Investigated process-safety incidents
  • Led an advanced process-control upgrade that saved about $2 million a year

Project Highlights

Route Activity Classification

Route Activity Classification

Ride with GPS

Is this route for road bikes, gravel, mountain bikes, hiking, or running? The model weighs several signals, like what the map says about the roads and trails along the way and how people actually use the route. It picks the right activity 95% of the time.

Trip to Route Matching

Trip to Route Matching

Ride with GPS

Figures out which routes a rider actually followed, even with noisy GPS. We use it to ask riders to review the routes they rode, which doubled completed route reviews.

Road Surface Prediction

Road Surface Prediction

Ride with GPS

Many roads on the map have no surface information, which matters a lot if you are on a road bike. This model predicts whether roads around the world are paved or unpaved, improving accuracy by more than 10%.

Nano-H3

Nano-H3

Fast H3 in one header

A single-header C++17 library that converts latitude/longitude to H3 cells. It matches H3 exactly and runs about twice as fast.

Autonomous Wheelchair

Autonomous Wheelchair

Curb Detection

A curb recognition and negotiation system that lets an autonomous wheelchair find a curb, approach it, and climb it.

Roof Shape Classification

Roof Shape Classification

Multimodal data fusion

A neural network that predicts roof shape from satellite images and LiDAR. I built a labeled roof dataset covering small to large cities and used it to evaluate the model.

Urgent Landing

Urgent Landing

UAS Rooftop Landing

An offline database of drone landing sites, each scored for risk to people and property. Flat roofs are found from LiDAR and images, then processed to pull out obstacle-free surfaces. A multi-goal planner picks the Pareto-optimal site.

Gas-fired Heaters

Gas-fired Heaters

Safety Systems

Safety instrumented systems (SIS) keep process units safe and running. I led SIS design and verification for four heaters through factory and site acceptance testing.

Media and Awards

Amazon re:MARS

Amazon re:MARS

Invited Speaker

Presented my work on finding urgent landing sites for small drones in cities, including ML that predicts roof shapes and Polylidar3D pulling landing sites off flat rooftops.

Swarm and Search AI

Swarm and Search AI

1st Prize Winner

A national DoD-sponsored competition to build algorithms that control a group of drones tracking a wildfire boundary. We took first prize ($27,000) and won the innovation award.

Into the Dataverse

Into the Dataverse

1st Prize Winner

A 3-day competition sponsored by the Joint Artificial Intelligence Center and NSIN. We presented an AI-enabled interface for recording maintenance logs and were awarded a $15,000 contract.

Awarded by First Sea Lord

Awarded by First Sea Lord

HMS Queen Elizabeth

We got the royal treatment as guests aboard the HMS Queen Elizabeth on Nov. 20, 2019, as winners of the Swarm and Search AI Challenge: 2019 Fire Hack. The First Sea Lord of the Royal Navy presented the award.

Last updated: