Lee Goymer Software engineer — backend & applied AI · former Amazon SDE · Brisbane
From a hard constraint to a working system.
I design and test software around concrete constraints, from unreliable upstream APIs to shared GPU memory and output-quality checks. I previously worked as an Amazon SDE and built a community platform's backend alongside a frontend partner. My current AI and systems projects are personal R&D, developed alongside part-time automotive-operations work.
Key results
- 2,771 / 2,835 on time ( 1 host, 1 card ; soak to 1 Sep 2026 ) GPU scheduler — deadline jobs on one shared card
How this was checked
Owner-reviewed 29 Sept 2026 · evidence held privately
One host, one card, and the consumer mix of that period; a batch-heavy window, not an interactive latency bound.
- 62.9% smaller ( n = 112 accepted files ; 22–29 Jul 2026 ) Video recompression — 112 accepted files; metric contract, not visual transparency
How this was checked
Owner-reviewed 7 Aug 2026 · evidence public
Fifty-four older records use the service's then-current minimum-floor constant because that field was not yet written per file; the result establishes contract enforcement, not perceptual transparency.
- ~40 public servers ( 41 in retained snapshot ; 28 Nov 2023 ) Community platform — external users, collaboration, and hand-off
How this was checked
Owner-reviewed 7 Aug 2026 · evidence public
Counts come from retained logs and a point-in-time snapshot; registered-user, total-vote, peak-concurrency, and availability figures are not claimed.
Selected projects
The work I would discuss first. Every project, with its state and limit, is on the Projects page.
| No. | Project | Type | State | Key evidence | Dwg no. |
|---|---|---|---|---|---|
| 1 |  Community platform Backend for a modded-Minecraft network Modpack voting with Stripe sponsorship and Discord entitlements, a stale-tolerant server list, and unattended publishing — run in production for a community of public game servers. | Case study2023–2025 | Retrospective | serve-last-known-good caching; atomic render-and-swap publishing Limit — Third-party evidence is what was retained | LG-006 |
| 2 |  GPU timeshare Multi-tenant scheduling for one RTX 3090 A constraint-aware scheduler for one RTX 3090 — a durable 24-hour plan, dispatch against physically free memory, and preemption that has to be earned by a committed beneficiary. | Case study2026 | Operational | 2,771 of 2,835 deadline jobs on time ( 1 host, 1 card ; soak to 1 Sep 2026 ) Limit — One card, and finite work still runs one at a time | LG-005 |
| 3 |  Video recompression under a quality contract A recompression service that binary-searches encoder quality against SSIMULACRA2, re-measures the full output independently, and refuses to ship below its contract. | Focused build2026 | Operational | 112 accepted files, 6.87 GB → 2.55 GB (62.9%), zero contract violations ( n = 112 accepted of 162 files ; 22–29 Jul 2026 ) Limit — A metric target is not visual transparency | LG-018 |
Now
What I’m working on — updated 29 September 2026. Full note →
I am preparing to return to a software team: selecting the strongest work from my independent projects and making its design decisions easier to inspect. My current technical focus is resource scheduling—what a planner can promise, and what still needs to be verified when a real process runs.
I am open to software engineering, research-software, and technical implementation roles with clear priorities and room for investigation, prototyping, and practical delivery.
Contact
Looking for hands-on software work.
Clear priorities, room to investigate, and a team to build with.