Bookmark ML System

A useful classifier begins with a usable dataset.

Bookmark collections become noisy quickly: labels drift, page text contains boilerplate, and a useful personal taxonomy rarely matches a generic public dataset. The system treats collection, cleaning, review, and modeling as one workflow rather than starting at model selection.

One workflow from browser capture to local inference.

• Bookmark Smart provides the browser-facing capture and organization experience.
• Page content is cleaned before keyword extraction so navigation, advertising, and repeated layout text contribute less noise.
• Bookmark Tools prepares JSONL training data, applies configurable heuristic labels, and supports human review through a lightweight Streamlit workflow.
• TF-IDF with Naive Bayes and logistic-regression baselines provides an interpretable starting point before considering more complex models.
• Saved model artifacts support repeatable batch prediction and local inference.

Evaluation is the next product feature.

The next pass will publish class distribution, split methodology, precision, recall, F1, confusion matrices, calibration, latency, and representative failure categories. Those results will be generated from a reviewed dataset rather than estimated for presentation.

• Prevent duplicates or near-duplicates from leaking across splits.
• Compare the heuristic bootstrap against learned baselines.
• Track performance by class, not aggregate accuracy alone.
• Measure the cost of human review and the value of corrected labels.

Privacy and limitations stay visible.

The local-first design keeps personal browsing data on the user’s machine, but it also means useful performance depends on each user’s examples and taxonomy. Sparse categories, changing sites, and ambiguous bookmarks are expected failure modes rather than edge cases to hide.