May 2026 – Present
Software Engineer
Lyft · Toronto
ML + analytics infra and data eng work.
00
About
Software Engineer
I build product, data, and ML systems with a bias toward clear execution.
Recent work spans full-stack product engineering, applied ML, and research projects in sports and decision systems.
01
Experience
Jobs, fellowships, and leadership roles that shaped how I work.
May 2026 – Present
Lyft · Toronto
ML + analytics infra and data eng work.
Sep 2025 – Nov 2025
Raremarq · Remote (Los Angeles)
Product eng work as eng hire #2.
May 2025 – Aug 2025
Lyft · Toronto
ML infra work in the marketplace org.
Sep 2025 – Present
Canada Basketball · Capstone
Homography computer vision research for sub-pixel player localization from broadcast video.
Sep 2024 – Dec 2024
Maple Leaf Sports & Entertainment · Toronto
Assorted full-stack, data, and product analytics work.
Jul 2024 – Aug 2024
Beats by Dr. Dre · Remote
Consumer behavior modelling for launch decisions.
May 2024 – Aug 2024
RBC Amplify · Toronto
Patent-pending recommendation engine for a gamified fantasy sports inspired financial education app.
Jan 2024 – May 2024
NBA
Youngest in cohort for executive-led training with fewer than 1% of 5,400 applicants accepted.
Mar 2022 – Mar 2024
Western Dev Society
Scaled a 700+ builder circle and partnered with 30+ organizations.
02
Writing
Working notes on product focus, reliable data, and keeping teams confident.
Turning broadcast tracking into a field of exploitable court space, then scoring screens, drives, passes and closeouts in the same units. Picks and drives open empty space, closeouts close it, and the number predicts the next three seconds of the possession.
OpenAI built a weird inference chip. I tried to rediscover its design choices from first principles: a simulator, an unbiased architecture search, adversarial tests that killed my NUMA story, and RTL that confirmed the one that survived.
How the Weldon Warriors placed 6th in Canada and top 3% worldwide at WorldQuant's IQC: 1,557 one-line primitive alphas, an overnight factory, and probing the scoring system before spending budget on it.
Every attention variant, RoPE, and a KV cache by hand on Candle, now running a real Qwen checkpoint, with predictions registered before the benchmarks came in.
03
Contact
Open to interesting product, data, and engineering conversations.
Best way to reach me is email.
lukeblommesteyn [at] gmail [dot] com