The Problem
Fine-tuning an open foundation model still looks more like research than engineering. Pull a checkpoint from Hugging Face, wire up a training loop, pick an accelerator, deal with CUDA or MPS, keep runs reproducible, ship the resulting adapter to somewhere it can actually be served: the stack is fragmented, the failure modes are unfriendly, and the tools assume you have the same Python environment as the person who wrote them.
Our Solution
TuneOps treats a tuning run as declarative infrastructure. You describe the model, the dataset, and the recipe; we run it on a local Kubernetes cluster on your laptop or on a remote GPU if needed. The same pipeline definition works regardless of where it runs. Runs are reproducible. Artifacts are tracked. The Desktop app is the first piece of this; more is coming.
Team
We’re a small team with backgrounds in ML infrastructure and developer tools. Get in touch if you’d like to talk.