What I've Built

One Saturday morning, I pulled up Claude Code and asked my 8-year-old, "What should we make with this?" After endless rounds of testing and iteration, her answer became Invisible Ink. This game is "exhibit A" of how far you can take a creative idea with agentic coding tools.

Underneath the fun-loving interface is a real system: real-time multiplayer on Firebase, API calls to Claude, 40 unit tests, deployed production security rules, a scoring overhaul that corrected the judge's anchoring bias, and continuous-improvement routines run by coding agents.

My day job is where I learned what it takes to trust AI output when the answers have to be right and the strategy has to serve the mission. The arc runs from writing the company's first AI responsible-use policy to training internal and external stakeholders on best practices, building a scalable research agent, developing a long-term strategy, and hardening data pipelines for 135k entities. Investing in AI has changed my career!

By the way, these aren't pipelines that simply say "find the answers." I bring in as much determinism as I can to scaffold the model's predictive nature: auditable logs, pre-processing that routes the level of effort, post-processing that assigns a confidence score, and more. The scaffolding is hardened by multiple rounds of detailed review and QA that buy confidence in the results later.

Aside protects something specific: where my voice goes. I liked Wispr Pro, but I didn't want their server to process it. So I built my own dictation app, open source, with audio that never leaves the machine: Python end to end, an on-device Whisper speech model, and a user-friendly SwiftUI interface.

It still shipped like a product: two tagged releases with DMG installers, and a test suite that grew from 77 to 149.

Temp Agency exists because a model's first answer is usually its most conventional one. It staffs a task with two research-backed specialist personas (a lead who drives the work and a contrasting lens who pressures it), so the output carries structured disagreement rather than a single smoothed voice. The workflow spans multiple models, not a single vendor's. Check it out for yourself!