AI Infrastructure
Engineering

I design, build and operate the infrastructure that runs AI systems — from the GPU and inference layer up to the product interface. Local and on-prem by default, tuned for latency, VRAM and cost.

Get in touch →What I Build

Infrastructure that runs AI — end to end

Most teams can call a model API. Fewer can run one well — on their own hardware, fast, cheap and reliable. I work across that whole stack: standing up model serving, squeezing latency and VRAM out of a GPU, building the MLOps and fine-tuning pipelines around it, and wiring in the agentic and retrieval layers that make it genuinely useful.

And because I'm also a design architect, the result isn't just a working backend — it ships with an interface and a developer experience people actually want to use.

What I Build

LLM Serving & Inference
llama.cpp / vLLM serving, GPU-tuned for latency and VRAM.
MLOps & Fine-tuning
QLoRA fine-tunes, eval harnesses and reproducible deployment.
Agentic Systems & RAG
Tool-using agents and citation-grounded retrieval, offline-capable.
AI-Integrated Web Apps
Next.js surfaces wired to model backends with streaming.
Design Systems & Architecture
Component systems, tokens and developer experience.
Platform & Developer Tooling
CLIs, TUIs, observability and latency/cost tuning.

Stack

Python · TypeScript · llama.cpp · vLLM · PyTorch · Transformers · CUDA · Unsloth / QLoRA · Next.js · Node · Docker · Linux · RAG / vector search · evaluation harnesses

Frequently Asked Questions

What does an AI infrastructure engineer do?

I build and operate the systems that run AI in production — model serving and inference, GPU and VRAM optimization, MLOps pipelines, fine-tuning, evaluation, and the agentic and retrieval layers on top. Effectively everything between the hardware and the product interface.

Can the models run on-prem or fully offline?

Yes — local and on-prem is the default. I run open-weight models on your own hardware (or a private VPC) using llama.cpp and vLLM, so there is no dependency on a third-party API and no data leaves your environment.

How do you keep inference fast and cheap?

Quantization, KV-cache and flash-attention tuning, right-sizing the model to the task, batching, and measuring latency and cost per request end-to-end. The goal is the smallest, fastest setup that still hits the quality bar.

Do you work remotely?

Yes. I'm based in Plymouth, UK and work remotely with teams anywhere. Engagements are project-based or contract.

LLM Serving & Inference →MLOps & Fine-tuning →Agentic Systems & RAG →AI Web Apps →AI in Plymouth →AI in Manchester →Journal →