Research Engineer
About UkisAI
UkisAI is an AI lab founded in 2023, with offices in San Francisco and Belgrade. We train frontier-quality models small enough to run on your own hardware. Our open Swift models use 40 to 60% fewer reasoning tokens than the models they're built on, and people have downloaded them millions of times. Ukis Foundation Models are next.
Our research is not academic for its own sake. It turns into the models we release, like Swift, and the people who download them are the ones who judge it.
About the role
We are looking for a Research Engineer to help make our models better and cheaper to run. You will work where research meets release: improving how our models are trained, quantized and evaluated, and turning promising ideas into measurable gains.
You will own experiments end to end, from forming a hypothesis to building the evaluation that proves or kills it. The work is empirical and fast moving, and the results show up in what we release.
In this role, you will
- Run experiments to improve our models, from inference-time methods to post-training with SFT, RL and on-policy distillation
- Quantize models and measure what each format costs in quality
- Build evaluations for the tasks we care about, and run them enough times to trust the result
- Curate, clean and generate the datasets that training depends on
- Investigate failures, like a model stuck in a reasoning loop, and turn what you learn into concrete improvements
You may be a good fit if you have
- A strong understanding of how modern language models are trained, fine-tuned, and evaluated
- Comfort with Python and the common ML tooling, and the ability to run experiments cleanly and reproducibly
- An empirical mindset: you trust evidence over intuition and build the measurement before drawing conclusions
- Curiosity about why models behave the way they do, and the patience to dig in
- Clear communication of methods and results to both technical and non-technical readers
Strong candidates may also have experience with
- Post-training open models with LoRA, SFT, RL or distillation
- Quantization formats and tools such as GGUF, AWQ, GPTQ or NVFP4
- Building evaluation suites and benchmarks for language models
- Data pipelines, synthetic data generation and dataset quality work
- Publishing, open-source contributions, or independent research projects
Logistics
This is a full-time, on-site role based in Belgrade. We review applications on a rolling basis.