I shape complex systems into tools people actually use.

An AI shortlisting tool for briefs nobody could parse

Clients arrived with messy, half‑formed descriptions of what they needed. The support team read each one by hand and guessed at search terms — slow, inconsistent, and impossible to scale with demand.

I built a tool that translated a raw client brief into structured queries against the marketplace and returned a ranked shortlist of sellers. The hard part wasn't the model; it was deciding how much interpretation to do on the client's behalf before the guessing became a liability.

Cut time‑to‑shortlist by roughly a third, and moved throughput off headcount.

A generative assistant for the worst part of hiring

Writing a good job listing is a skill most buyers don't have, and a vague listing poisons every match downstream. Rather than adding more form fields, I led a generative assistant that turned a rough brief into a polished, marketplace‑ready listing — collapsing a multi‑step creation flow into something closer to a conversation.

Simplified the highest‑drop‑off workflow in the funnel, and improved the quality of what entered the matching system.

Three compliance services that unblocked enterprise hiring

Enterprise buyers couldn't transact on a marketplace their legal and procurement teams hadn't cleared. The blocker was never the quality of the work — it was unsigned agreements, unverified identities, and contractor‑misclassification exposure that carries genuine legal and financial liability.

I shipped three services, each integrating a specialist third party behind a single entry point on the freelancer's profile:

• Background checks Criminal and adverse‑media screening through Checkr, with freelancer consent built into the flow and a result the client reviews and acts on themselves.
• Legal documents DocuSign‑backed agreements an admin uploads once and the team sends per freelancer, tracked from sent through signed or expired.
• Worker classification audits Independent‑contractor risk assessments through Greenlight, triggered by a screening questionnaire and returned as a clear risk determination.

The hard part was coherence. Three vendors, three unrelated processes, three different timelines — and a client who should only ever have to learn one. All three share an entry point, a status model, and a place in the Pro subscription tiers. Deciding what to abstract away and what the client genuinely had to see was most of the work; a compliance product that hides the wrong detail is worse than no product at all.

All three are live, and publicly documented in Fiverr's help centre.

AI governance starts with the data underneath it

Enterprise and government customers can't govern AI without first governing their data. Most arrive asking about model oversight and find the harder problem sitting underneath it: they don't know what they hold, where it lives, or who can reach it. An AI system grounded on that estate inherits every one of those gaps, and inherits them silently.

I own product strategy and roadmap for AI governance and data lifecycle workflows — the parts that turn a sprawling, largely unstructured data estate into something a records manager or a regulator can actually interrogate. The customers are public sector, financial services and critical infrastructure, so every decision the product makes has to stay defensible years after whoever made it has moved on. That constraint shapes the work more than any feature request does.

Alongside it: a prioritisation framework balancing customer value, strategic fit, quality and delivery complexity, now used across the organisation — and a design system paired with an AI‑enabled prototyping workflow.

Roughly halved the time from customer feedback to a working UI component.