The work, and what it did.

Thousands of SKUs needed manual image checks against product name, team, colour and description. Slow, inconsistent, expensive — and legacy AI let subtly wrong logos slip through while breaking downstream pipelines with non-JSON output.

One production-hardened prompt per attribute, each with a strict output contract, a clear mismatch taxonomy, and a reasoning field explaining every flag. Shipped as the org's canonical prompt library.

Hundreds of campaign assets a week across four leagues. Manual review couldn't hold, and existing prompts weren't built for bulk or for non-technical users.

The full prompt layer of an internal tool: upload up to 1,000 images, pick the audit type, get downloadable CSV results with Slack notifications — decomposed per league for precision.

Multiple LLM workflows running at once, but no standard way to compare models, justify upgrades, or track cost and quality over time. Every use case triggered a bespoke, unreproducible evaluation.

A HELM-inspired, cross-use-case evaluation framework on DocETL pipelines — structured multi-model comparison, socialised internally as the active tool for model-migration decisions.

Prompts written like QA checklists caused chats to loop back to intro after resolving, mis-classified complaints, and drove escalation cost up.

Re-architected instruction design for six agent types, first-message intent routing, and conversation-lifecycle logic that ends gracefully. Diagnosed and killed the root-cause chat-loop bug.