At the desk
Building a private agentic AI finance project. More when there's something to show.
Software engineer and general funny man. I build cloud-based Mac environments that customers run on Apple Silicon at scale: continuous integration and delivery (CI/CD), mobile device management (MDM), artificial intelligence and machine learning (AI/ML), virtual desktop infrastructure (VDI), and computer-using agents (CUA). I write the code, not just the architecture. After hours I ship iOS apps, macOS tools, and open-source Swift.
I'm a software engineer. I've shipped software at Capital One, Uber, Robinhood, Apple, and Amazon Web Services (AWS). These days I design and build the CI/CD, MDM, AI/ML, VDI, and CUA systems customers use to run fleets of Macs on Apple Silicon.
After hours I build open-source Swift projects — napkin, SFSymbolsKit, and Spooktacular, a macOS virtual machine (VM) manager — and I write.
Right now I'm building distributed agentic systems:
fleets of AI agents that run on real hardware and produce real
outcomes, not demos. I care about local-first software,
readable typography on small screens, clear
READMEs, and software that works for the person
using it.
An open-source macOS VM manager for Apple Silicon: ephemeral runner pools, warm-pool scrub validation, Kubernetes orchestration. One box delivers two Macs' worth of CI capacity. 424 tests, MIT-licensed.
A Swift 6.2 framework for building apps as a tree of isolated, composable units — Uber's RIBs pattern, rebuilt for Swift Concurrency. I reimplemented it in project after project, so I published it.
A small Swift package for working with SFSymbols. Extensions on String, UIImage, and NSImage; the full symbol catalog is generated by a Python script in the repo.
Building a private agentic AI finance project. More when there's something to show.
Halfway through Nixon's The Real War. Reading it slowly, on purpose.
TOPS, Griselda, Part Time, Yves Tumor, Ghostface Killah, Vince Staples, Big Body Bes, Meyhem Lauren, BADBADNOTGOOD, and Kanye West on rotation.
Mentoring a few engineers and remembering how much I like it. Trying out new bits at open mics — stand-up is harder than distributed systems. Baking cookies and making ice cream, and slowly getting good at both.
An agent is a non-deterministic system, so grading it by the artifact at the end means comparing noise. Track the path instead — the sequence of tool calls it made, and the state of the workspace at each one.
An agent that does a thing on its own is a product — that's the whole word. The model inside it is a commodity you swap; the part you can't commoditize is the person deciding what it should do.
A follow-up to "What an AI Agent Actually Is" — how the parts compose into a single agent, how several agents compose into a system, and why your tool set is the part that's actually yours.
Use the links below. Reach out about a project, a problem, or music I should hear.