About the Role
At WYLE, we are building an AI that runs the parts of your life you should not have to think about, so you get your time and attention back. Not another app. A Personal Operating System for people whose lives are too full to manage by hand.
It has to work reliably in one real person’s life, over years, in the actual mess of a real day. That is a hard engineering problem before it is anything else.
This role owns one part of it: WYLE will not stay a wrapper around someone else’s models. We are building toward smaller, faster models we run ourselves. You own that work, end to end.
You have fine-tuned real models (SFT, LoRA or QLoRA, preference tuning) and can show the result. Inspectable weights, benchmarks, or an eval someone can check. If you have never shipped a model another engineer could look at, this is not your role.
Key Responsibilities
- You quantize with intent (GPTQ, AWQ, GGUF, FP8) and own the accuracy-versus-latency trade-off rather than guessing at it.
- You have run self-hosted inference in production (vLLM, TGI, SGLang, or Triton), made it faster or cheaper, and can explain exactly how.
- You own the full loop, from data to training to quantization to serving. You are not a research scientist who hands off a checkpoint, and your AI experience is not just calling an API.
- Bonus: on-device or edge inference, distillation, and low latency on ordinary consumer hardware.
How to Apply
More jobs at get9to5jobs.com