Last updated: 2026-05-09
This is a page in the spirit of nownownow.com: what’s in focus right now, without plans for a decade.
Work
AI Automation Engineer (contract), TaxDome — since January 2026.
I design and deliver end-to-end LLM/agent workflows in Python and TypeScript: LangGraph for orchestration, LangChain for core components, n8n for glue. Direct calls to OpenAI / Anthropic / Gemini APIs where fine-grained token and latency control is needed.
Current focus:
- Eval systems. Prompt regression suites via LangSmith traces, structured outputs, and golden datasets — to prevent agent quality degradation as prompts and models evolve.
- Retrieval pipelines. Chunking, embeddings, hybrid search, reranking over internal documents and CRM data. Services on FastAPI, agents — on the consumer side.
- AI MVP end-to-end. Backend + lightweight frontend + AWS infrastructure in Docker — to quickly validate hypotheses and cut time-to-value.
- Continuous reassessment of new models, agentic frameworks, and tooling (Claude Code, Cursor) — what works as a daily driver for the team.
Personal Projects
- artka.dev v2 — final rollout. All four EPICs from the spec
2026-05-02-llm-citable-blog-design.mdare shipped: unified JSON-LD@graph, retrieval frontmatter (summary/keywords/faq),<Tldr>/<Faq>/<Compare>/<Definition>/<KeyTakeaways>MDX components, tag archives and related posts via Jaccard. Now polishing Phase 5: courses on top of the same content model, progress via PostgreSQL + Better-Auth, on-demand PNG certificates, and per-course RSS. - Claude Code Guide → course. A series of 14 posts migrated into a full-fledged
course with interactive exercises (
<ExerciseCheck>,<CodeChallenge>), auto-progress via dwell-time + IntersectionObserver, and a certificate at 100% completion. - MCP tooling. Personal GitNexus (code graph) + llm-wiki as a daily driver for agents. Experimenting with tool-design boundaries: what to give agents first-class, what goes in reference docs, what goes in hooks.
Recently Closed
- Blog SEO audit — two waves of fixes. Identity refresh via JSON-LD graph
(Person with real CV,
sameAs, raster avatar, and Organization logo), title and description tuned to AI/backend profile, BreadcrumbList + WebPage / CollectionPage on every landing, JSON Feed 1.1 alongside RSS, per-page OG for all landings. - Phase 4 dev ergonomics. View transitions on title/cards, Plausible (DNT-aware), split sitemap by locale, reading time, course player scaffold.
- Phase 3 retrieval & SEO. Pagefind ⌘K, RelatedPosts via Jaccard, per-post OG via Satori build hook, Buttondown newsletter, Giscus.
- Phase 2 reading. ReadingProgress, footnotes as side-notes on ≥1280 px,
rehype-autolink-headings with
#anchors, Shiki dual-theme (github-light/github-dark-dimmed) with[data-theme]swap. - Phase 1 design system. Paper/sienna palette, dark variant, ThemeToggle with no-flash bootstrap, lockup in mono.
What’s Next
- Postmortems on production incidents in agent systems. Turn accumulated raw material from 9RED Wallet and current work into public writeups on tool design, eval loops, and guardrails (ETA: December 2026 – January 2027).
- Second course. Still in outline stage, direction: “Production-ready agent loops: harness, eval, guardrails.” If there’s a topic you’d like to see — let me know.
- Open-source several MCP servers once I stabilize the API.
If you’d like to discuss — write to a@artka.dev or Telegram.