I build things that span the whole stack — from the copper traces on a circuit board to petabyte-scale data pipelines, and lately from novel transformer-based architectures to the laser-cut enclosures that house them. Over twenty-odd years I’ve been a graphics researcher, a platform engineer at Google, a founder soldering my own boards at 2am, and most recently someone who teaches other people how large language models actually work under the hood. If there’s a thread, it’s a deep fascination with making powerful, complicated technology feel approachable — whether the person on the other side is a Pixar lighting artist, a Google Maps routing algorithm, a roomful of grad students, or me chasing an I²C timing bug with an oscilloscope.
I read Computer Science at Cambridge and started out writing compilers for graphics hardware — and I’ve never lost the compiler-writer’s habit of wanting to understand a system all the way down. It’s the thing that lets me pick up a new language or domain quickly, because underneath they mostly rhyme.
University of Washington — Teaching how LLMs actually work · 2025–present
I teach at the University of Washington iSchool, which keeps me current and honest in a way that shipping alone never quite does.
- Introduction to AI. (IMT598L/K)
- Building LLMs (IMT526A): a full, hands-on in-depth tour through constructing and implementing large language models — the architectures, the current state of the technology, the very real risks, and a lot of coursework. You don’t really understand a transformer until you’ve built one and watched it fail interestingly, retrained it, and created something new.
Infinite Variations — Founder, electrons to AI · 2024–present
I’m a long-time electronic-music producer, and I’d been quietly irritated that hardware sequencers make you work one track at a time — you can’t see how the parts fit together. So I founded Infinite Variations to build the sequencer I wanted, and did it as a one-person team: electrons to embedded firmware to platform services to a neural network, all in one box.
- Designed the six-layer PCB and hand-placed the 0402 components (sub-millimetre — I wrote my own reflow temperature-control firmware to do it), and brought the board up over JTAG/SWD.
- Wrote real-time firmware on a QP/QXK microkernel holding sub-2ms MIDI latency, with a Python multiprocess system orchestrating audio, UI and hardware over custom IPC — plus a browser-based hardware emulator so I could develop without the physical unit.
- Built a transformer from scratch in PyTorch for controllable generative music, designed to run on-device on resource-constrained Arm hardware.
- Even designed the switched-mode power supply — and a custom signal-injection transformer to characterise it.
- Concept to production-ready hardware in under a year, by overlapping cycles: firmware while PCBs fabbed, architecture while units validated.
Google — Senior Software Engineer, Maps · 2014–2023
Nine years building large-scale developer infrastructure and ML systems for Google Maps.
- Detected road closures from traffic patterns at 95%+ accuracy by combining black-box ML with old-fashioned queueing-theory statistics — predictions that fed directly into routing for millions of people every day.
- Ran petabyte-scale processing on Borg, high-throughput pub/sub pipelines ingesting millions of sensor readings an hour, and multi-tenant privacy boundaries.
- Built the pipelines that ingested COVID quarantine-zone data into the maps during the pandemic.
- Mentored engineers and interns, sat on the hiring committee, and taught CS fundamentals to high-schoolers through Google’s summer institute.
Pixar — Senior RenderMan Developer · 2008–2014
Relocated to the US (and later became a citizen) to join the small team that reinvented RenderMan as a plugin-extensible Monte Carlo path tracer.
- Built online, self-tuning importance sampling that learns where light comes from as it renders rather than relying on a fixed analytical model — roughly 10× faster on hard lighting.
- Architected the geometric area-light system that became the foundation for lighting in modern RenderMan.
- Wrote a SIMD JIT compiler for the shading language, and published the work at SIGGRAPH 2012.
Framestore — Senior Developer · 2005–2008
Led the fur and grooming system — millions of individually simulated hairs per character — that won a Technical Achievement Academy Award (and a BAFTA) for The Golden Compass.
- I’m still quietly delighted there’s an Oscar attached, somewhere, to the problem of making a polar bear look furry.
- It succeeded because we sat in dailies with the artists, documented in their language, and obsessed over workflow. Good tools meet people where they are — a lesson that shaped everything after.
In the open
A lot of my tinkering ends up open source: I’ve contributed to Neovim, I maintain the zsh grammar for tree-sitter, and I’ve sent patches as far as the Linux kernel. I also keep a small constellation of developer tools — dotfile syncing, a modular zsh framework, AI-in-the-shell experiments, and tooling for managing MCP servers across LLMs — most of it on docs.georgeharker.com.
Off the clock
I live in Seattle. I make electronic music (the through-line behind half of this page), cook, take photographs, and spend as much time as I can hiking and camping in the mountains — usually plotting the next thing to build on the drive home.
Work with me
I work best on the bridge between users and highly complex problems — not just as a translator, but as an innovator. The thread through everything above is exactly that: taking what people actually need — artists, drivers, students, musicians, fellow engineers — and turning it into systems that are powerful underneath and genuinely usable on top. I’m at my best when the path isn’t mapped yet, when the problem is still fuzzy and the right approach has to be invented rather than looked up — and working across hardware, platforms and AI lets me reason about a problem from the silicon to the user and back, and find the leverage others miss.
I’m open to opportunities that involve real technical challenge — analysing and driving development in areas of uncertainty, particularly across Product Design, AI, and Electronics. After a stretch building solo at Infinite Variations, I’m especially keen to do that as part of a strong team, where the hardest problems only get sharper for being argued over.
If that sounds like your kind of problem, I’d love to talk. Reach me at george AT georgeharker DOT com — or wander through georgeharker.com, GitHub and LinkedIn.