Three ways to work together.
An audit when something's broken, a build when something's missing, and deployment when the model works but isn't a service yet. Pricing is transparent because surprises are for birthdays.
The AI Audit
Your agent or RAG pipeline is confidently wrong in production and you can't work out why. That's a measurement problem before it's a model problem. I find the failure modes, rank them by impact, and fix them.
- Retrieval quality assessment
- Chunking & embedding review
- Prompt, context-ordering and grounding analysis
- Agent tool-calling & routing failures
- Written report with reproducible failure cases
- Fixes implemented, not just listed
The Build
Embedded engineering on one or two hard problems — an LLM or vision system taken from prototype to something measurable and maintainable, or the full stack around it. Within 48 hours of the first call you get a written scope and architecture.
- LLM agents & RAG pipelines with citation grounding
- Computer vision, including CPU-only deployment
- The full stack around the model — Rails, FastAPI, React, Postgres
- Evaluation harness and regression suite you keep
- Weekly written checkpoints and running code
- Handover documentation — you're never locked in to me
Model to Production
A model in a notebook isn't a product — and it probably doesn't need a GPU. I turn trained models into fast, monitored, documented services, usually on CPU, which makes the inference bill a fraction of the quote you were given.
- Clean REST API with input validation
- Dockerised and portable
- ONNX conversion and quantisation where it helps
- Measured latency and throughput, before and after
- Model registry, versioning and drift monitoring
- CI/CD pipeline
Four phases, no theatre.
Scope
A call, then within 48 hours a written scope and architecture: components, risks, milestones, and how we'll measure success.
Baseline
Before changing anything, I measure what you have. Without a baseline, every improvement afterwards is an opinion.
Build
Weekly written checkpoints and working code you can run yourself. No status meetings, no progress theatre.
Hand over
Documentation, the evaluation harness, and a handover doc — so your team can maintain it without me.
Common
questions.
Are you open to full-time roles?
Yes — actively. I'm London-based and open to AI/ML engineer, data scientist or senior backend roles, including sponsored positions. Freelance work runs alongside that, not instead of it.
How quickly can you start?
Audits usually start within a week or two. Larger builds depend on what's already running — I take a small number of projects at a time so each gets real attention.
Do I need a GPU budget?
Often not. I served an ArcFace ResNet-50 through ONNX Runtime on CPU-only hardware at 32ms per frame. Plenty of production inference is a CPU problem that gets quoted as a GPU one — I'd rather cut your bill than pad it.
What if AI is the wrong answer?
Then I'll say so, early, and tell you what I'd build instead. A retrieval problem solved with a well-tuned index beats an LLM that hallucinates confidently, and it's cheaper to run.
Who owns the code?
You do. Work-for-hire terms in the contract, clean documented code, and a handover document so nothing depends on my continued availability. Prior open-source work stays open source.
Can you prove the numbers on this site?
Yes. Every figure traces to a primary source — the dissertation and coursework reports, the marked transcript, or a role I can put you in touch with. Ask and I'll send the document.
Ready to
start something?
Tell me what you're building and what's blocking it, and I'll tell you honestly whether I'm the right person for it.
Get in touch