Home / Building / AI Lab

The AI Lab.

Experiments in every state: shipped, half-built, abandoned, and still just ideas. Most AI pages show demos; this one shows process, because the abandoned ones taught me the most.

building

Portfolio Built with AI Coding Agents

The question: Can AI coding agents (Claude Code, Hermes delegation) architect and build a production portfolio site — Remix RR7 + Mantine + Sanity + MongoDB + Cloudinary — with a human steering design decisions and copy?

What happened: In progress. AI handled scaffolding, component structure, route architecture, and Sanity schema generation surprisingly well. The design system, copy, and information architecture still need human taste. Biggest learning: AI is an excellent junior developer but a poor design director. Currently iterating on the Failure Museum and AI Lab pages.

See it live

shipped

Hermes: Autonomous Daily Briefing Agent

The question: Could an LLM agent run autonomously on cron, track my Mac activity, capture window titles and browser tabs, and deliver structured daily briefings without human intervention?

What happened: Shipped and running daily. 9AM check-in, 7PM briefing, 9PM health check, Sun weekly review. Tracks per-app minutes, parses VS Code project names, browser tab titles, Slack/Teams conversation names. The surprise: the hardest part was parsing window titles meaningfully (every app formats them differently), not the LLM integration. Built with a VPS agent orchestrating over SSH to Mac.

shipped

AI Blog Pipeline: 38 Technical Posts

The question: Could an AI-assisted pipeline produce publication-ready technical blog posts at scale — from drafting to Sanity CMS to Vercel deploy — while maintaining a consistent, authentic voice?

What happened: Shipped. 38 posts published covering TypeScript, MongoDB, WebAssembly, AI trends, Bun, React patterns, and more. AI drafts reach ~70% quality. The last 30% — voice, the right anecdotes, the 'I tried this and here's what broke' honesty — can't be automated. Established a hard rule: an AI draft never publishes without a human edit pass.

shipped

DocPilot's AI Copilot

The question: Could an AI layer brief a doctor's morning, summarise any chart in seconds, and read practice analytics in plain language, without ever diagnosing or prescribing?

What happened: It shipped, live inside DocPilot. The hard part wasn't the model — it was the restraint. Every summary stays a summary. The deterministic product works without AI; AI only ever adds, never blocks. Built for Indian clinics where a prescription in ~30 seconds means a doctor leaves on time.

See it live

Some experiments graduate. Some go to the museum.

the lab. mind the unfinished ideas.