continual learning

Learning log

Paper notes, experiments, tutorials and things I learned this week — my gradient updates, in public.

Accuracy isn't enough: calibrating a tool-selection model

Fine-tuning Qwen3.5-2B to pick tools took accuracy from 50.6% to 100% — but the number I cared about was ECE, 0.220 → 0.002.

Why dense-only retrieval fails on identifiers and jargon, and how fusing sparse and dense results before a cross-encoder fixes it.

What actually breaks production chatbots: indirect injection, system-prompt extraction and excessive agency.

Designing DocuMind AI, my first hot-tier estimate was ~10.5 TB. The right answer was ~21 GB — here's the mistake.