Writing
Notes on building agents, AI products, and the tools I make along the way. Mostly things I’ve built, decisions I’ve made, and what I learned from using them.
2026
Qty 10- Building custom apps for my girlfriend's oral boards4 min
My girlfriend is preparing for pediatric dental oral boards. I've been building her custom apps for studying, interview practice, and the trip around the exam, plus a dental game and a local model experiment whose first evaluation gave me a useful negative result.
- One plugin repo for all my MCP servers3 min
Five hosted MCP connections share portable plugin definitions and generated client manifests. Keeping the instructions aligned with the live tools is the next maintenance problem.
- Practicing system design with an AI on the whiteboard4 min
A shared Excalidraw canvas gives my agent the design I am practicing. Semantic edits, saved versions, and feedback on the actual board keep the discussion grounded in my attempt.
- The grocery MCP server behind my Ask DoorDash pitch4 min
A Kroger MCP server for my own groceries became the prototype behind my Ask DoorDash pitch. Its current design separates product search, selection, and cart writes, with explicit recovery when a purchase action is uncertain.
- One source for the ways I present my work3 min
A résumé, a CV, a website, and a chat interface need different presentations. A private JSON source, redacted public API, and shared renderer keep the facts connected without exposing every field.
- Returning to projects I thought I was done with5 min
Revisiting old repositories with agents has become part of my workflow. The useful changes preserve old data, consolidate study tools, and repair checks. Each needs evidence at the level of the behavior it restores.
- An agent needs somewhere to work4 min
Tools, persistent state, clients, shared workspaces, and status displays give an agent's work a life beyond a single chat connection. The design question is which part owns each durable record.
- Building my own interview practice environment4 min
My interview projects range from solution archives to shared whiteboards. The common question is how to make assistance improve the practice itself, while keeping problem statements, saved solutions, and feedback separate.
- From meal ideas to a grocery cart4 min
Recipes, product matching, shared lists, and pantry inventory are different parts of the same grocery problem. Reviewable drafts and durable records make each handoff easier to check and recover.
- A website for every adventure4 min
Travel planning has become one of my most useful agentic workflows. I build a website for each adventure, with one trip model behind its maps, timeline, packing list, and read-only assistant tools.
Writing elsewhere
Articles I coauthored with the DoorDash engineering team.
DoorDash Engineering Blog ·
A platform for building and evolving agents
The shared platform behind Ask DoorDash: agent orchestration, MCP tools, durable state, and the tradeoffs in deciding what to standardize.
DoorDash Engineering Blog ·
Building DoorDash Assistant: an engineering overview
How intelligence, evaluation, platform, and user experience fit together in a conversational shopping product.