Smart Experiment Management & Meta-Analysis for Visdom

Visdom is great for watching training in real time, but once you have dozens of experiments, keeping track of what hyperparameters you used in each one becomes a mess. This project adds an experiment tracking layer to Visdom a SQLite-backed metadata store where every run carries structured hyperparameters, metrics, and tags. I am building a search system so you can query runs with simple syntax like "lr less than 0.01 and accuracy above 90", interactive parallel coordinates plots to visualize the hyperparameter space, and Optuna integration so Visdom can suggest what to try next based on past results. Everything runs locally with zero external dependencies, keeping Visdom's self-hosted philosophy intact. I have already contributed 6 PRs to the Visdom codebase covering backend bug fixes, frontend improvements, test fixes, and Python 3.12+ compatibility, and I have deep familiarity with both the Tornado server and the React frontend.