Senior Member of Technical Staff (Inference Acceleration )
Description
What you will do:
- Own inference performance for the models we deploy: kernels, batching, cache management, quantization, speculative decoding, and the profiling that tells you which of those actually matters.
- Cut latency and cost without giving up quality, and prove it with measurements rather than estimates.
- Work with the routing and embedded teams so that serving decisions reflect what the hardware can really sustain, including the thermal and power envelope.
- Build and evolve the agent harness: how tasks are decomposed, how tools and models are called, how traces are captured, replayed, and analysed.
- Work on self-improvement honestly: which experience is worth keeping, how to judge it without self-flattery, and how to turn it into a policy change that generalises instead of memorising.
- Treat evaluation as part of the loop — a harness change that is not measured is not an improvement, and a score that went up for the wrong reason is a regression.
- Keep the framework usable: the people running experiments on top of it should not have to read your internals to do their job.
What we are looking for:
- A strong academic background in systems, ML, or a closely related area. Publications are a strong plus, especially on efficient inference, agents, or self-improving systems.
- We equally welcome candidates whose excellence shows through other evidence: an inference stack you built, competition results, or a framework with real users.
- Solid systems engineering: profiling, low-level performance work, and the judgement to know when optimisation has stopped paying.
- Comfort with the uncomfortable part of RSI: knowing that "the agent got better" is a claim that needs evidence, not an observation.
HOW TO APPLY
Apply via LinkedIn. All applications are handled in strict confidence. Only shortlisted candidates will be contacted.
J&L Apex Advisory Pte. Ltd. | EA No. 26S3529 | EA Personnel No. R1329267
Skills
Sign in to read and join the discussion.