About
Andrew Nordstrom
I build infrastructure for software that makes decisions and takes action, along with the control layers that make it trustworthy.
In practice, that means recommender systems people can inspect and steer, and infrastructure that keeps agent-driven development coordinated and auditable. Research Software Engineer at That Recommender Systems Lab in Boulder, Colorado.
How I Work
My default mode is taking a system from concept to production and staying with it. Corgi started as an early lab idea. I proposed moving it from Mastodon to Bluesky, built it, and now operate the live feed. CARAWAI started as a proof of concept and is now an end-to-end platform. Staying through operations is where most of my design opinions come from.
When I change a system with real users on it, I stage the risk. My ProPublica work for POPROX is deliberately staged: the new source stays isolated from the default candidate pool and existing subscriber flows until the downstream path is validated. The foundation lands without changing current production behavior.
Research questions change faster than systems can be rewritten, so I put the volatility behind contracts. Corgi's epoch model and registry-driven scoring let ranking policy evolve without touching the pipeline. The stable parts of a system should be the ones people rely on: the contracts, the audit trail, the coordination rules.
I try the standard tool first and measure. When off-the-shelf NLP couldn't extract what I needed, I replaced it with custom LLM pipelines and a confidence-weighted merge. When my own concurrent workstreams kept colliding, I built the control plane that now coordinates them. The tool follows the problem.
I work inside a research lab, so my output is other people's input. Three PhD dissertations depend on software I build and maintain. That's why I'm precise about what results show: a synthetic evaluation is not a community outcome, and a single trial is not a general result. A claim that's exactly the right size is one a collaborator can build on.
Location
Boulder, Colorado
Education
University of Colorado Boulder
BS Mechanical Engineering
Minor in Data Science
Current Role
Research Software Engineer
That Recommender Systems Lab
Stack
TypeScript, Python, PostgreSQL, Redis, AWS, Docker, AT Protocol, Chromium extensions
Background
I've been building applied ML systems since 2018. The first was a wildfire detector: a CNN on a Raspberry Pi, a motorized zoom lens sweeping the horizon, and a text to first responders with GPS coordinates when confidence crossed 95%. I trained it against the conditions that fool a camera, like sunsets and storm clouds, because a detector nobody trusts is just noise. It won Best in Fair and Intel's Excellence in Computer Science award two years running.
The next system went from reporting to deciding. My LiDAR irrigation prototype scored plant health from moisture, pH, imaging, and elevation mapping, then opened servo-actuated floodgates zone by zone. In its one controlled trial it used 1.21 gallons of water per day against 4.41 for flood irrigation, roughly 73% less. One prototype, one trial. I treat that as evidence the system worked, not proof it generalizes.
Mechanical engineering at CU Boulder added the physical half: modeling, fabrication, and end-to-end design. On a five-person team I led mechanical design for a humidifier prototype that took first place in its category at the 2023 IDE Design Expo.
For three years I mentored a high school researcher building CNN-based skin lesion screening for Android. She presented at Regeneron ISEF in back-to-back years, earned a $10,000 scholarship, and was named a 2024 National STEM Champion. Coaching the full research arc, from model design through defensible experiments and technical writing, is still some of the hardest engineering communication I've done.
Today the systems are bigger and other people's work depends on them. At That Recommender Systems Lab, led by Professor Robin Burke, I lead engineering on Corgi and CARAWAI and contribute to POPROX. Three PhD dissertations run on software I build and maintain. Outside the lab, I run my own projects on an engineering control plane where agents handle implementation and deterministic infrastructure handles ownership, review, and recovery.
The scale keeps changing. The question doesn't. All of it is software acting on the world, and all of it is only as good as a person's ability to check what it did and why.
I build the system that acts. Then I build the layer that lets you check its work.