Pagefind's stemming warning was a red herring — Japanese search broke on segmentation
Pagefind warned about missing stemming for Japanese and Chinese on every build. The real defect was a query and index tokenizer mismatch returning wrong pages.
All posts
git gc recovered 8.6 MB while I was investigating a 0.8 MB problem in git history
Chasing 1.6 MB of PNG blobs in git history found a repo that had never been packed. git gc recovered 8.6 MB; the rewrite I skipped was worth 0.8 MB.
Astro share buttons that pick their targets by language — a sticky rail on desktop, pills on mobile
Share buttons in Astro built from intent URLs: a different target set per locale, a sticky icon rail on wide screens, and the endpoint drift found on the way.
Playwright as a measuring instrument — the i18n defect no static check can reach
Measuring translated SVG labels against their boxes with Playwright, and the two bugs that made the check report 37 pages clean while measuring nothing.
An Astro table of contents that keeps your place — scroll-spy, a shared rail, and dvh on mobile
Follow-the-reader scrolling, a shared rail, and the ::details-content box that silently ate a mobile max-height in an Astro table of contents.
Expressive Code in an Astro blog — filename tabs, file icons, and 97 unreadable fences
Adopting Expressive Code in Astro 7: filename tabs, per-theme file icons, a Sätteri processor trap, and 97 fences Shiki was reading as a language name.
How a chain of failures produced the LLM (part 2) — from attention to the scaling revolution and alignment
From context to the finished LLM: the limits of RNNs, the birth of attention, the Transformer, GPT-3 scaling, RLHF alignment, and the era of efficiency.
How a chain of failures produced the LLM (part 1) — from the limits of maths to the geometry of meaning
The LLM as a by-product of unrelated problems: Turing on computability, Shannon on information, the perceptron, backpropagation, GPUs, and Word2Vec.
RAG vs fine-tuning for KB search — what fine-tuning actually learns, and when pure IR wins
Whether fine-tuning can replace RAG for knowledge base search: what fine-tuning actually learns, when pure IR or long context wins, and what each costs.
The HTTP QUERY method (RFC 10008) implemented in Python — how it differs from GET and POST
Implementing RFC 10008 HTTP QUERY in Python with Starlette: safe, idempotent, cacheable requests with a structured body, and how it complements GraphQL.
Deterministic AI agent tests with AIMock — mocking LLMs, vector DBs, MCP, and A2A on one port
AIMock mocks the whole AI stack on one port (LLM, vector DB, MCP, A2A), so an agent test suite runs in under three seconds with identical output every time.
A Claude Code /logo skill built from five wrong turns — SVG paths, regex bounds, a wrong kanji
Redesigning this blog's logo and code font with an AI agent: a broken SVG letterform, a regex bounding box, a wrong kanji, and the /logo skill it produced.