About

Usman Bukhari

I am a senior AI engineer. Bukhari Consulting is my training practice.

Delivery record

I build production LLM pipelines and agent systems: retrieval over large document estates, structured extraction from unstructured records, and agent workflows that act on real systems rather than demonstrations.

I have delivered that work in national healthcare with the NHS, in aerospace airworthiness compliance with Rolls-Royce, and in regulated financial services and insurance with Ardonagh. These are environments where output is reviewed, decisions are recorded, and being wrong has consequences that reach beyond a sprint.

I am still building. The training is a description of my current practice, not a course written once and repeated.

Approach to teaching

My sessions are worked, not watched. Teams bring their own documents, tasks and codebases, and we apply the tools to those directly. I teach nothing that I have not used in anger on a real deliverable.

I give failure modes as much time as capabilities. Participants learn where these tools are confidently wrong, because that is where the cost sits — a fabricated figure in a client report, an invented API in a merged pull request, a silent breaking change that passes review.

I deliberately keep product names out of the material. Tools are renamed and replaced constantly; the judgement transfers.

Why regulated domains shape it

Building where an auditor may ask how a result was produced changes what good practice means. You need to know what data went into a tool, what was checked before the output was relied upon, and who signed it off.

That discipline is the most transferable part of what I teach. It is why responsible adoption — what can safely be put in, how output is verified, and what a sensible internal AI policy says — is core material rather than an appendix.

Elsewhere

I publish selected code and experiments on GitHub.

Book a conversation