Founding AI Researcher
Lucebox designs hardware and software together to make fast local AI something you own. Its Lucebox Zero 495 is a private, plug-and-play AI inference workstation shipped with an open source inference engine and tuned models, sold to individual developers and to teams and companies.
Projects
About Lucebox
Lucebox (AGMI AI, Inc., designed in Italy) builds the Lucebox Zero 495, a local AI workstation with an AMD Ryzen AI MAX+ PRO 495, up to 192 GB unified memory and a 32 GB Radeon AI PRO R9700 GPU. It runs Ubuntu Server with an Apache 2.0 open source inference engine, custom kernels and speculative decoding, so large open models such as DeepSeek V4 Flash and Qwen 3.8 27B run locally with no per-token cost. Setup is done from a browser, and the machine exposes OpenAI and Anthropic compatible APIs for tools like Claude Code and Codex. Lucebox sells to individual developers through an online shop, and to teams and companies through a Business tier with installation, model tuning and engineer support. It publishes engineering benchmarks on its blog and models on Hugging Face. It is backed by Y Combinator, Vento Ventures and AngelInvest Ventures.
Skills
Candidate Availability
Required and preferred rules are kept separate and reflect the wording in the original posting.
About the Role
You will be the first researcher, leading work on making open models better over time on hardware people own. You will take open problems from idea to result with experiments you design and run yourself, post-train and fine-tune open models end to end on our own hardware, and build the evaluation measurements that show whether a change improved a model. You will set the research agenda with the founders and see what works ship into the open source engine.
Requirements
- PhD in machine learning, finished or close to it, or an equivalent record with first-author papers at venues like NeurIPS, ICML, ICLR, ACL or EMNLP
- Hands-on experience post-training models and understanding GPU hour usage
- Fluency in PyTorch and ability to read kernels when speed is a problem
- Conviction about running research on machines people own
- Nice to have: released research code or models used by others
Responsibilities
- Take open research problems from idea to result with self-designed experiments
- Post-train and fine-tune open models end to end on own hardware
- Build evaluation measurements to determine whether a change improved a model
- Set the research agenda with the founders
Benefits
- Equity of 1% to 2%
Hiring Process
Intro interview (30 minutes with a founder), technical interview on machine learning and your research, behavioural interview, technical project on a small research question, final decision within a week of the project.