I turn personal friction into public tools and dig into the why behind the libraries I use. Side experiments: @thecuriousts.

Featured and Cooking did not come from a master plan. They grow the way real craft does: a spark on X, friction from another domain, a personal limit, a season of life you cannot negotiate away — then the need to make something so the constraint loses a little power.

One thing triggers the next. Connections show up after you move. Exploration under real limits beats over-planning everything early and shipping nothing.

If you are waiting for perfect conditions: start with the friction in front of you.
The list below is what that looks like when you keep going.

What’s unique · stack · why look it up. The chain above is how they arrived.

Not a job board scraper — a self-guarded Tauri reactor: live X search + xAI fit/prep, SQLite ledger, secrets in Rust, pause guards so agents stop instead of thrashing.

Learn — productized agent autonomy with honest HITL; MVU desktop; credential boundaries that survive real use.

Not another dotfiles dump — a kernel/plugin shell (zsh · bash · fish) with PATH contract, environment presets, and an ab/av/at verification cockpit for reviewing agent work.

Learn — portable shell product design; plugin boundaries; terminal UX for AI-era operators.

Profile-based Arch bootstrap that audits and hardens itself into an ML/AI-ready workstation — installer as ongoing security product, not a one-shot script.

Includes eye-comfort — not random theming: vision-science circadian phases (warm nights, soft days, dawn/dusk), solar bounds from latitude, and location / culture / season overlays (Tamil tinai · perum · nāḻikai) so the palette respects place, calendar, and eyes on long sessions.

Learn — systems product thinking; self-audit loops; theming as ergonomics product, not aesthetic cosplay.

Skills extracted from real projects (orchestration, security, multi-agent work) — not generic prompt packs.

Learn — skills as reusable engineering assets; high-signal packaging over chat trivia.

One optional Zod model → per-consumer required fields at runtime, with OpenAPI interop — static schemas stop where multi-tenant APIs start.

Learn — validation as product API; schema craft that scales without duplication.

Elm-style MVU for C (program core, Cmd/Effect, composition) — predictable architecture on pure C. Powers premflow (daily-driver CLI) as real dogfood.

Learn — functional architecture under systems constraints; frameworks proven by a product you use every day.

Privacy-first harvest of your own agent transcripts → datasets, skills, gold exemplars — training material without shipping chats to the cloud.

Learn — agent tooling as data product; local-first pipelines.

Tiny trainable LSTM + minimal ReAct so the code teaches the idea — education, not production AI infra.

Learn — first-principles LLM mechanics you can inspect and re-run in an afternoon.

Tamil classical grammar & poem prosody (metres, segmentation) — rare domain software with a Rust/WASM hot path and modern web UI.

Learn — cultural-computational product design; WASM for linguistic structure.

Portfolio as a live product: agent-oriented Q&A, document views, automated PDF CV — presence engineered, not a static README clone.

Learn — Next.js product polish; CV/PDF pipelines; personal brand as maintainable software.

Still under heat — and a little semi-planned: as more work can be delegated to agent systems, it helps to see overall progress without killing natural inception (spark → friction → next link).

Shipped to learn — live surfaces, not daily drivers. Useful when you want the idea under your fingers before productizing.

Two habits that outlast any single project: orchestrate under constraint, and when the calendar slips, ship more than was asked — deeper work, not a thin late apology.

Core idea (thesis → 2026): an orchestrator that profiles, predicts, and acts under constraint — not “battery tips.” Same shape as today’s AI stack problem: where inference runs (cloud · edge · on-device NPU), what data leaves the machine, how agents spend energy and attention. Intelligence that respects cost and locality compounds; blanket centralization taxes humans.

• Articles on X — culture, health, tech, policy; careful long-form when stakes feel real, not take-farming.
• Books — Rust from First Principles Companion — language nuances, interconnectedness, non-linear learning (with Grok). From shelf-life.

Latest first · first-publish year.