How I work

I spent a decade on observability — the tooling that tells you what your system is actually doing when it is under load and nobody is watching. That work rewires how you build. You stop trusting the happy path.

The regulated-product years

At Overjet I owned the frontend of an intraoral imaging capture system that went through FDA 510(k) clearance. Regulated software has a reputation for slowing people down. What it really does is force you to be specific: about what the software claims, about what happens when it fails, about what evidence you have that it works. I found I liked it.

The problem I am proudest of solving there was mundane and hated. Dental clinics have bad networks. Capture sessions were dying mid-exam when the backend blinked, and a lost session means a patient sits back down in the chair. I converted the app to a PWA, built real offline resilience into the capture path, and the session finalization rate settled at 99.5%. Then we took it from one pilot clinic to 145.

Why on-device

In 2023 at New Relic I shipped two LLM features into production for enterprise customers, back when that still meant arguing about whether it was a good idea. Since then the interesting constraint moved. Cloud inference is a solved commodity problem. Getting a real model to run well on a phone — inside a memory budget, without draining the battery, without shipping the user's voice to anybody — is not.

So I built things. Earshot puts Whisper-grade transcription on both iOS and Android from one Kotlin core, at 37% less peak memory and better word error rate than the platform recognizer. Sitewinder runs three models at once on an Android phone with the radio off. Pencil Count read 406,769 scanned ballots with a local vision model and landed within 1.4% of the official count.

None of these are demos. They are the parts of the on-device problem you only learn by hitting them.

How I am on a team

I work end to end and I go where the risk is. In practice that means I am usually the person who picks up the reliability problem everyone has been routing around, and the person who says the schedule is slipping while there is still time to do something about it.

I mentor deliberately, and lately most of that has been about AI-assisted development — not the demo version, the version where you have opinions about what to hand a model and what to keep. I use agentic tooling daily and I have views on where it earns its keep.

Outside the editor

I am in Portland. I lift, I garden, I run a small radio observatory on a Raspberry Pi that watches aircraft and ship traffic and flags anything odd in the spectrum. I lost about seventy-five pounds and wrote watch apps to keep myself honest about it, which is why this site has a sibling called Code to Muscle.