medgemma
New
› medgemma
────────────────────────
status active
runtime mlx · apple silicon
target on-device inference
────────────────────────
$ pip install medgemma
Upload data, generate training sets, fine-tune Gemma 3n via QLoRA on spot GPUs, export production-ready mobile models. 90–97% cheaper than SageMaker, under 48 hours end-to-end.
I'm an AI engineer focused on the space where machine learning meets real hardware constraints. I build tools that make sophisticated models practical — running privately on a laptop, deploying cheaply on mobile devices, or training efficiently on spot GPU instances.
My work sits at the intersection of model fine-tuning, on-device inference, and developer tooling. I'm particularly drawn to problems where the default answer is "use a cloud API" — and finding a better one.
When I'm not building I'm reading about quantization techniques, new edge inference frameworks, or thinking about how small teams can own and operate AI systems without the enterprise price tag.
Have a project in mind, want to collaborate, or just want to say hello? I'd love to hear from you.
Currently open to freelance projects, interesting collaborations, and full-time roles in AI engineering.
Arte, scientia, et arte aedificandi
in nova aetate consilium et analysisem.
Dispatches on AI, edge systems, and the ideas behind the work — exploring the overlap between technology, craft, and the act of building. Published on Substack as KANUNI.