Michelle Chen

Michelle Chen is Group Product Manager for Cloudflare’s AI Platform and helped start AI Gateway, which sits between applications and model APIs to cache responses, limit requests, track usage, and route around provider failures. Her work connects access to models with the practical decisions developers face when running them: which provider to use, how to manage spending, and what happens when a request fails.

From WARP to AI Gateway

Michelle worked on WARP clients as a Cloudflare product-management intern in 2019. She later became chief of staff in the company’s Emerging Technology and Incubation group, a role that included helping develop new products. In her account of starting AI Gateway, she describes drawing on that product-management background when the group needed help preparing its new AI offerings for launch.

AI Gateway entered beta in September 2023. Michelle coauthored the announcement with Yo’av Moshe and Meaghan Choi. Developers using a supported provider could adopt it by changing the destination URL of an API call while keeping their existing payload structure. That placed Cloudflare in the request path without requiring developers to rebuild their application around a different model interface.

The initial features addressed specific costs of operating an AI application. Cached responses could answer repeated requests without another paid model call. Rate limits could constrain spending and excessive use. Analytics brought request counts, latency, errors, and costs across providers into one place. A separate universal endpoint accepted provider-specific payloads and let developers configure retries and fallback models.

Michelle explained fallback with a concrete pairing: an application could try OpenAI first, then send a request to Workers AI if the first provider failed. This could keep a provider outage from becoming an application outage, although the fallback meant using a different model. Her contribution was helping establish the product that made these shared operational needs available to developers.

Making model choice usable

In the same interview, Michelle distinguished Workers AI, which runs models on Cloudflare-managed GPUs, from AI Gateway, which manages requests to those models or external providers. She connected access to open models with reducing dependence on one company’s decisions. The combination gave developers somewhere to run models and a way to route requests among alternatives.

Her subsequent coauthored work follows the development of both products. An August 2025 announcement with Abhishek Kankani and Mia Malden introduced access to models and dynamic routing through one endpoint. In August 2026, Michelle and Ming Lu announced the unification of Workers AI and AI Gateway into a single AI control plane, bringing observability, billing, and routing across Cloudflare-hosted inference and external providers. These developments extended the original proxy into a common interface for using and managing multiple model services.

Bringing live web information into the request path

Michelle coauthored the October 2026 Web Search API launch with Sam Else and Gabriel Massadas. The integration addressed a different limitation: a model’s training cannot supply newly published documentation or other changing information. Search could retrieve fresh, structured web snippets for an agent’s context, including documentation for developer tools released after its training cutoff.

Search requests used AI Gateway’s existing logs, access controls, and credit balance; developers could also bring their own provider keys. The launch began with Ceramic.ai, Exa, and Linkup. It tied participation to concrete crawling requirements: providers had to identify their crawlers, respect robots.txt, meet Cloudflare’s verified-bot requirements, and return links to the crawled material. The announcement presented those requirements as a way to give site owners a say in how their content was used. Native server tools were still forthcoming, distinct from the search interfaces available at launch.

Michelle’s work also extends into models designed for decisions inside software. She coauthored the October 2026 introduction of Clef and Clef-flash with Alex Reneau and Kevin Flansburg. The announcement summary describes open-source decision models hosted on Workers AI for classification and agentic workflows, alongside a reinforcement-learning platform for fine-tuning them with developers’ own data. This adds task-specific model adaptation to the inference and request-management products she has helped develop.