Cloudflare's engineering teams now use AI for 93% of their development work. This adoption, reached in under a year, is supported by a platform handling 241 billion tokens monthly through its AI Gateway. The system's output is concrete: weekly merge requests increased from 5,600 to over 8,700. The architecture, built entirely on Cloudflare's own products, offers a blueprint for internal AI deployment.

The technical core is AI Gateway, which routes requests, manages costs, and enforces a zero data retention policy. It processed 20.18 million requests last month. For inference, the stack uses both external providers and Cloudflare's serverless Workers AI, which runs on global GPUs to avoid cross-cloud latency. Their internal Kimi K2.5 model, handling 7 billion tokens daily, addresses cost and is used for lighter tasks like documentation review.

Engineers access everything through a single proxy Worker. Authentication flows through Cloudflare Access, issuing a signed JWT. The proxy, a Hono application, forwards requests to AI Gateway with an anonymized UUID, stripping personal identifiers. Configuration is managed as code, compiled and validated at deployment.

Key to scaling was solving token bloat from Model Context Protocol (MCP) tools. Their 'Code Mode' collapses many tools into two generic endpoints, controlling context window size. Enforcement is automated; an AI Code Reviewer agent, integrated into every GitLab pipeline, assesses each merge request and delegates analysis to specialized sub-agents for security, performance, and compliance.

The platform demonstrates dogfooding at scale. It tackles engineering problems like schema management and attribution without PII, suggesting that the primary advantage for Cloudflare is a stack that operates entirely on its own network.

Source: Webpronews