From the Sunset District, San Francisco
I’m Tony, an SRE in San Francisco. Nights and weekends, I build software with a focus on privacy.
See what I’m building01 / About
I live in the Sunset District with my family and two Shiba Inus who have perfected the art of treating every request as a suggestion.
I’m training for an ultramarathon, and all the running gives me time to think up projects and work through how I might build them with AI. I spend nights and weekends turning those ideas into software.
I build from the same starting point: local-first, minimal dependencies, no tracking.
A native Bitwarden and Vaultwarden client for iOS, iPadOS, and macOS. No third-party code in the runtime.
I reimplemented the crypto in Swift and check it against RFC vectors and weekly differential tests with the upstream C reference. Vault data stays ciphertext at rest. On macOS, an ssh-agent signs git operations behind Touch ID.
A browser engine built from scratch in Rust, with 23 crates covering HTML parsing, layout, and GPU rendering.
A screenshot-reference test corpus and per-subsystem maturity docs say where compatibility ends. I’d rather publish the honest maturity table than the demo. The repository is private for now; I’m happy to walk through it.
An MCP server that gives AI models scoped access to Kubernetes clusters. Secrets are redacted by default, and destructive operations require human confirmation.
KUBECONFIG never leaves the local process. The project is still in progress.
Presence-first messaging for small, trusted networks. Inspired by ICQ, with a buddy list and numeric UINs.
Accounts are invite-only and never ask for a phone number or your address book. Use the hosted version, or self-host it in about fifteen minutes. The project is still in progress.
03 / Process
Building with AI has become an experiment in how I work. I’m learning how to describe an idea, give the models useful context, and check what comes back. I’ve tried pairing development agents with QA agents and having models interview each other to surface questions I hadn’t thought to ask.
The models are starting to feel more distinct to me. Some respond well to detailed instructions and frequent course corrections. Others do better with a clear goal and room to figure out the approach. I’m learning when to be specific, when to leave things open, and how to adjust when the same guidance works well with one model but gets in the way of another.
Those lessons sometimes mean starting over. More than once, I’ve gone back to a project and rebuilt it from scratch because something I learned along the way changed how I wanted to approach it. The next attempt starts with a better understanding of both the software and how to build it.
04 / Contact
If you want to talk about a project or compare notes on building with AI, email me.