Built in a hurry. Used for the thesis. Highlight → meaning in the popup → pick a better word without leaving Energy Efficiency as an Orchestration Service. Wiktionary and Tamil Virtual University stay in the popup. Grok and Grokipedia when you want the subject and more context, not just a gloss — they open in a new tab so the extension does not pay model API bills. Current Chrome extension format (Manifest V3). Tap for meaning; go deeper only when you ask.

archy is a thin top controller plus domain jobs (Eagle + satellites) — not a pile of dotfiles. Thin host, then optional machine-learning or security profiles, evidence bundles, inventory, weekly scans. groxy talks to a Grok agent over the ACP, or pings you on X. eye-comfort: body-clock lighting + latitude + Tamil tinai and the நாழிகைnāḻikai · 24-minute clock. After Omarchy resets the status bar, personal-tweaks put the Heading chip back.

Kernel versus plugin on zsh, bash, and fish: a PATH contract, plus ab / av / at panes to review agent work. An ontology skill so agents stop moving files across the kernel boundary.

One optional Zod schema. Each API consumer names which fields it requires at run time. OpenAPI $ref in, Zod out. Multi-tenant APIs where a static required: list dies.

Scars, not best-practices. Formal-first SKILL.md. EVA emptiness = a prior when the map is missing (do not pretend you already know). Fission then fusion of context. Cognitive-load dual track: a human’s working memory versus an agent’s parallel tools. Extracted after the fighting stopped — not designed in a vacuum.

Harvest Cursor + Grok Build + Kilo + Cline on disk. Raw transcripts, harness dumps, and agent dotfiles stay local. Only the distill step talks to a language model — and only tidied, extracted turns, not the bulk session. Then Rhai workflow scripts. No bundled dataset.

A tiny character LSTM (next-token net), a ReAct loop (reason, then act, then observe), memory, a self-critique data factory, and a typed workflow — the agent-improves-itself flywheel you can hold in your head. prototype-it-to-explain-itself

Retrieve-then-answer was already in xAI Collections. Upload persona and CV — xAI splits the text and builds the vectors (I noticed them appear on upload; it was not a loud feature while X was still yelling “build your own retrieve-then-answer stack”). /qa only searches (documents/search) and asks Grok. No chunker, no embed model, no Pinecone or Postgres-vector store of my own. Local qa-index.json is the fallback when that reactor is off. /focus.

• Thesis PDF — hosted full text · devprofile#32
• Uppsala DiVA — record · FULLTEXT01 (often blocked off-campus)
• IEEE CloudCom 2015 — EEaaS paper
• Wiley 2017 — profiling paper
• On X — long arc post · /focus write-up

Prefer hosted PDF when DiVA fails.

• Report PDF — ML face-recognition (2011)
• On X — delayed work, shipped more

Uppsala ML course — assignment ran late; extended into real implementation (finished solo). Same muscle as today.

• Zod nullish chaining — PR #1702 (2022)
• react-intl — babel-plugin-react-intl-messages-generator (2017)
• Kent C. Dodds — bookshelf #116 · #184 · testing-react-apps #39 · #42
• React Boilerplate — PR #1355 (merged 2016) · #1364 · #1167

Uppsala years: room to wander. Pro years: low time budget, honest stubs — not a victory lap.