AI Spaceship

by Mugilan Ragupathi

I build AI agents — and the harnesses, tools, and guardrails that make them work in production. Each project below is a live demo you can try, with its source and docs.

Start with the AI Agent Harness series to see how a language model is turned into a tool-using agent, or work through the AI Concepts lessons to understand how a model reads text in the first place.

Projects

  • ShopBot

    A customer-support agent that resolves order questions end to end.

    Does
    Looks up orders, checks live tracking, and issues refunds — in plain conversation, no forms or menus.
    Built
    A hand-written agent loop (no framework): the model calls typed, schema-validated tools and reads results back until it answers. Refunds require human approval, tool failures are returned to the model to recover from, and the Anthropic call runs in a FastAPI backend so the key never reaches the browser.
    Next.jsFastAPIAnthropic APISQLiteagent looptool use
  • AI Concepts

    Interactive lessons on how a language model actually reads text.

    Does
    Teaches tokenisation, embeddings, attention, output logits and context windows in short lessons — a worked example, then questions, with hearts, XP and a streak.
    Built
    Content is data, not UI: a lesson is a list of typed screens, so a new chapter is authored rather than coded, and the routes, lesson map and sitemap all derive from it. Counting questions compute their answer from the real o200k_base vocabulary at runtime instead of storing one, so a question can't drift out of sync with the tokeniser. Every lesson prerenders to its own shareable URL, and it runs entirely in the browser — no backend, no keys.
    Next.jsReact 19gpt-tokenizero200k_basecontent as datano backend