About | Agention

Twenty years, four industries

I've been a software engineer for over 20 years, working on a variety of projects and applicatons across healthtech, fintech, media, and climate tech. Recent stops: Sylvera (carbon credit ratings), Doxy.me (healthtech, the V2 video call platform serving 100k+ daily interactions), StepStone (staff engineer across 7 front-end teams). Before that: contract work at News UK, Barclays, Caplin, and others.

The through-line is architectural rigour on real production systems, event-driven pipelines, Domain-Driven Design, cloud infrastructure, and being hands-on while doing it. I've written 20+ technical articles, founded a startup, mentored a lot of engineers, and led department-wide workshops on system design, performance, WebRTC, and React internals.

Starting Agention

In February 2025 I started Agention as an open-source project alongside contract work. What began as a TypeScript agent library grew into a small stack: Marshall (coding harness), Signal (Qwen fine-tunes), the AP quant series, and a fork of llama.cpp with adaptive speculative decoding tuned for AMD's Strix Halo APU. All of it built on a single Strix Halo box.

Faster than expected

The trajectory has been faster than I expected. The adaptive-speculation patch was widely adopted in the Strix Halo community, and the feedback that came with it, that the hardware had too many competing forks already, is what led me to start Halo Box, a community fork aimed at being the fastest inference engine for Strix Halo rather than one more fork among several. The AP quants have 100k+ downloads on HuggingFace. Signal trended on the image-text-to-text page, and official Qwen releases, and has been independently validated. Marshall benchmarks at about 50% token savings versus Codex CLI on the same tasks.

Coming to ML from systems

My background isn't in ML, it's in production systems. That intuition transfers more than I expected. Inference stacks, samplers, quantization tradeoffs, and speculative decoding are all engineering problems in the end, and 20 years of production systems work turns out to be a decent foundation for reasoning about them. I'm learning the ML side in public, one shipped artifact at a time.

Outside of it

Outside all of this I've run six ultramarathons and climbed in the Himalayas and Alaska. Same posture: sustained focus on hard problems, respect for the terrain.

Get in touch