AI Influencer Matching at Creator.co
The creator-matching page in Creator.co's brand dashboard, where London, the platform's AI agent, returns ranked creators. I built the frontend and the server pieces it needed, on a global team.
Creator.co is an influencer-marketing platform. Brands find creators, run campaigns and talk to them in-app, and London is the platform's AI agent. The matching page is where a brand describes what it wants in a sentence and gets ranked creators back, or switches to manual filters and browses. Between November 2025 and March 2026 I built most of its frontend, in React and TypeScript, on a global team with product managers, QA and backend engineers.
This is the one project on this site that wasn't mine alone, and that's the point of including it. Product set the scope. QA tested every change with screen recordings and I answered with recordings of the fixes. I built on an API handler a colleague had written rather than adding my own. The frontend was mine; on the server I owned the pieces the page needed, and the rest of the search backend was the team's.
The creator card. The new card from the design, with lazy-loaded media, truncation, a responsive layout and a fallback for creators with no media, with unit tests from the start.
The page itself. Connecting the search service to the page and opening the new route; agent mode and manual mode kept cleanly apart; filters, sort and platform tabs; bulk save across platforms; search state that survives leaving the page, so a brand can open a creator and come back to the same results.
Performance. Manual mode had been slow enough to feel broken, so I cached profile images and spaced the requests. A hundred-plus results with three images each stuttered, so the list only renders the cards on screen. Stale results no longer flash while filters or sort change.
Server-side pieces. Wiring the MCP search API to the UI, infinite scroll that stays compatible with the old search, bulk save, and one small change I'd point to over the rest: the search was also requiring the words "influencer" or "creator" in a profile bio, which most creators never write, so it was discarding good matches that topic, hashtag and interest matching had already found. Removing that check was the single biggest improvement to results.
Instrumentation. The AI-search usage event had never fired in production, so nobody knew how the feature was being used. I fixed it and added events for switching mode and applying filters.
Separately, I built the reply suggestions in brand chat end to end: a new intent on the agent, drawn from the campaign's stage, deliverables and payment status; the gateway call with a timeout; and the frontend, which shows suggestions above the input, including when a chat is already waiting on a reply, and tracks which ones get used.
The rule I'm most attached to is in code, not the prompt. Any suggestion that depends on payment is dropped even if the model ignores the instruction, and the model is stopped from inventing deadlines or promising scope. A suggestion a brand clicks without reading is a commitment, so the floor is enforced, not requested.
Most of the value in a feature like this isn't the AI call. It's the hundred decisions around it: what to show while results load, what a mode switch should reset, which events to fire so the team can see what's happening, and what the model is never allowed to say. That's the work I'm proudest of here.
Live at creator.co.