A practice in building AI systems

I build what
AI works in.

Environments to act in. Memory to build on.
Systems people can understand and shape.

Currently building Magpie and researching Fusion Compute.

Explore the work GitHub

The environment around intelligence

Give agents an environment where code becomes action.

Puppys

EducationQueen Mary University of LondonMSc Artificial Intelligence · 2022

ExperienceFirst engineer, PuppyAgent TechTechnical leadership · MiraclePlus Fall 2023

Planning a Bay Area visitSeptember–October 2026

01 / Current work

Two questions.
Built into systems.

One explores how intelligence runs.
The other, how we put it to use.

Fusion ComputeResearch study / 01

One pod · two GPUs

GPU AState αNVIDIA L4

GPU BState βNVIDIA L4

Decode layerNext Q / K / V

250 / 250 audited stepsRecorded experiment · Qwen2.5–1.5B

Distributed inferenceResearch

Fusion Compute

Can inference keep going
while state stays split?

I’m exploring Transformer inference without gathering all execution state in one place. The recorded experiment runs a decode layer on two L4 GPUs, then continues into the next layer’s Q/K/V projections.

Read the technical briefing 中文

What the experiment establishes

250/250 audited steps passed with no hidden AllGather inside the declared boundary: one full decode layer through the next layer’s Q/K/V projections, on two NVIDIA L4 GPUs in one pod.

This record does not establish continuous multi-layer inference, cross-provider execution, or cryptographic privacy.

MagpieIllustrative workflow

A real conversation

For your review

Something said.
Something to build on.

Connected to the source conversation

KeepEditLeave out

Conversation → a reviewable artifact

Human × AIProduct in development

Magpie

Good conversations leave
something worth keeping.

Magpie turns useful moments from professional conversations into cards you can review, edit, and reuse. I’m building a workflow that keeps the person involved in deciding what matters.

Explore Magpie

A small design-partner pilot

For independent consultants, coaches, and advisors: bring one working conversation, then review the captured moments together.

The question is simple: does it produce something you would actually keep, edit, or use in a follow-up? The review workflow and customer-value validation are still in progress.

02 / The work behind the work

A continuing
line of inquiry.

Agents need somewhere to act. Workflows need structure. Knowledge needs to stay connected. Each project carries a question into the next.

From the first commit
to the first team.

I like the part where an idea has to become a working system.

At PuppyAgent Tech, a MiraclePlus Fall 2023 company, I was the first engineer and de facto technical lead. I took on the initial architecture, product delivery, engineering hiring, and leadership of a team of up to five.

Before that, I earned an MSc in Artificial Intelligence from Queen Mary University of London in 2022.

See what I build

Open to conversations

Bay Area · Sep–Oct 2026

Working on a
related question?

I’m planning time in San Francisco, Berkeley, and Silicon Valley. I’d like to meet AI builders, research communities, and people interested in trying Magpie.

AI infrastructure. Human–AI tools.
Early-stage founder communities.

Open to a longer stay for the right project and community.

Find me on GitHub

I build whatAI works in.

Two questions.Built into systems.