all projects

project 21 / 31

MLOps (Self-Hosted W&B)

Fully offline Weights & Biases instance for training and tuning - real-time monitoring of sweeps, gradients and loss curves with zero data leaving the network.

#mlops#weights & biases#self-hosted#machine learning#hyperparameter tuning#data privacy#python

About this project

Training and tuning models offline is a vibe - but the monitoring is where the real magic happens. Watching a complex experiment unfold in real-time, catching bugs before they burn through compute, discovering problems you wouldn't have seen otherwise.

Setup: a self-hosted Weights & Biases instance, completely offline.

Air-gapped security: all simulation data and model weights stay strictly within own infrastructure. Full visibility: real-time logging of system metrics, gradients and loss curves without a single packet leaving the local network.

Customized sweeps: rather than W&B's built-in sweeps, the logic lives in the code itself - deeper customization tailored to axial flux machine parameters and simulation constraints. The parallel coordinates plot becomes the roadmap through high-dimensional data toward the best-performing model.

hosted on fly.io