AI Enablement
How I help teams adopt LLM tooling, and where the evidence comes from. Every item here comes from work I did, mostly at Mastercard.
Scope of my code claim: I read and review code. I ship with agentic tooling. I do not hand-write production code, and from-scratch coding tests are my weakest area. I am strongest at architecture, review and running delivery. If your role needs someone writing production code line by line, I am the wrong hire.
What I Have Delivered
Agentic content review for a Mastercard product team
Recent work at Mastercard includes AI review skills, an orchestrator and an eval harness I designed and built with the MDES Manager design and content team. 27 skills delivered in a versioned bundle, about 75 defined, from an audit of 166 screens. Detectors run in parallel, mechanical rules run in code, facts are guarded twice and a separate grader scores the result. It is a proof of concept with light testing, and it is not in production. The design is on the Agentic Content Review page.
LLM-driven rewrite of a 380-page guide
At Mastercard I rewrote the MDES Manager user guide into about 200 pages of DITA XML, using LLM tooling, with standardised terminology, tone and formatting. I worked section by section with product managers, product and marketing. Details on the Mastercard page.
RAG-ready content repository
I designed and built the content repository behind Mastercard's conversational-AI proof of concept: 14 delivery guides in GitHub. Each guide has self-contained, intent-tagged sections and Q&A pairs, chunked and terminology-controlled for retrieval accuracy.
Precision on the claim: I built the content and its architecture. I did not build the assistant itself.
LLM-assisted review every release cycle
I review 30 to 40 screens per release against Mastercard's content standards, with an LLM doing the first pass.
Applied computer vision, built with agentic tooling
I built a tool that turns survey video into classified traffic counts, measured against a manual count. See Traffic Counting.
How I Work With Agentic Tooling
- Tools: Claude Code, the Claude API, MCP and agent skills, every working day
- Design pattern: the model advises and vetoes, deterministic code executes. Every LLM call goes through one gateway, every verdict is logged and graded
- Earned authority: a model gets more autonomy only after its logged verdicts show it has earned it. I use this on my own autonomous trading agent, described on Independent Builds
- Review: someone has to understand what the code does and answer for it. That is the part of the job I keep
Where My Judgment Goes
AI makes drafting cheap. These parts still need a person who is accountable for them:
- Standards reconciliation: turning inconsistent guideline documents into instructions an agent can apply, and deciding the ambiguous cases
- False-positive control: an agent that flags the wrong things at scale is worse than no agent. Define the quality bar and test it on real work
- Scoping: deciding which use cases carry real value and which are demos
- Adoption: fitting the tool into the team's existing workflow so it stays in use
I did the same split in my platform audits. Tools produced raw metrics. The value was calling the false positives and finding the root causes. See Case Studies.
Earlier Evidence of Driving Adoption
- Folk Digital: built the delivery process from nothing and got a team to adopt it: version control, code review, sprints, Jira and Confluence. New Relic went in as standard on every build
- Klevu: onboarded and documented an AI search product for merchants and developers, and hosted developer webinars
- StarkWare: docs-as-code, analytics-led prioritisation and automation experiments against a constantly shifting protocol roadmap
What I Do Not Claim
- A production AI deployment. The Mastercard agent work above is a proof of concept
- Hands-on production coding
- Formal compliance or AML expertise. My payments exposure is on the security and third-party side