Software engineer building automation systems, AI tooling, and web platforms.

Designed the full system: agent graph, tool contracts, retrieval layer, audit trails. The old process was a spreadsheet queue. Replaced it with a state machine that agents operate inside of.

Most of the work went into making it legible to non-technical reviewers. Every action is traceable and tied back to the source document.

Wrote custom migration plugins to normalize 14 content types across three source databases. Editorial kept publishing in the old system until cutover, and diffs were synced continuously.

After launch: render cache strategy, image derivatives on a CDN, and search reindexed to Elasticsearch with boosts editors can tune themselves.

Started as "automate the forms handling" and turned into a full integration project. Legacy SOAP endpoints, a GDS design-system front end, and a review board that needed human checkpoints at specific stages.

Ended up as a small set of independent workflows instead of one big pipeline. Each has its own SLA, failure routing, and owner.

Built a labelled eval set from real tickets before writing any retrieval code. That's what told us when hybrid scoring was actually helping vs just adding noise.

Shipped it behind the same internal UI the support team already used. The old search box just got better, no training needed.