BUILDING SYSTEMSTHAT THINK.

Binary frames, hidden information, and partial observations make it easy for a tracker to sound more certain than the evidence allows.

Page-start WebSocket instrumentation, an in-browser MessagePack decoder, raw-first capture, event inference, and a local dashboard that distinguishes exact knowledge from conservative estimates.

Makes uncertain game state auditable: decoded evidence survives, unknown cards remain unknown, and risk is explained instead of guessed.

Most coding assistants collapse the learning loop by revealing the solution. Useful mentorship must improve the next decision without taking it away.

A Next.js and TypeScript workspace with deterministic review scheduling, explainable weak-topic queues, public LeetCode activity import, and optional Gemini-assisted grading.

Keeps the scheduling core deterministic and inspectable while using AI to improve ordering and feedback—not replace the mastery model.

Dataset preparation becomes slow and error-prone when annotation controls and export formats hide the geometry users are actually editing.

Direct-manipulation boxes, reusable color-coded classes, keyboard workflows, and bidirectional conversion between normalized YOLO coordinates and pixel-based Pascal VOC bounds.

Moves an image folder toward training-ready labels through a transparent project format that remains interoperable with downstream ML tooling.

Bulk deletion is irreversible, rate-limited, and easy to target incorrectly. A safe interface has to make scope and consequences visible first.

A dependency-free Node server and vanilla interface that compose text, regex, link, attachment, date, and exact-ID filters into previewable, stoppable jobs.

Turns risky cleanup into an observable pipeline with preview counts, progress metrics, bounded logs, and explicit stop controls.