About the role
You explore how much useful intelligence fits into an affordable personal computer, and you pursue only the improvements that survive honest evaluation. Clear negative results are a real contribution here, because they stop the rest of the company spending a quarter on something that will not work.
The work
Research compression, distillation, sparsity, architecture or adaptive computation. Compare against strong baselines on both model capability and actual device behavior. Work with inference and hardware engineers to test whether theoretical savings become user-visible improvements.
What good looks like
In your first 90 days, reproduce a baseline and deliver one tested hypothesis with quality, latency, memory and energy results.
Evidence we look for
Bring strong mathematical and experimental foundations, with a doctorate or equivalent research evidence. Clear negative results and reproducible code are valuable contributions.
What we need to see
- Strong mathematical and experimental foundations, with a doctorate or equivalent research evidence
- Reproducible code published alongside your results, not on request
- You report negative results clearly rather than reframing them
- You evaluate against something honest rather than against a favourable baseline
Nice to have
- Model compression, distillation, or sparsity specifically
- Architecture search or efficient architectures
- Open-weights contributions
The exercise
Explain an efficiency result that failed after deployment and propose an experiment to identify the hidden cost.
Where and how we work
In the office together five days a week, in any of these cities. Remote-friendly around your family, arranged one person at a time.