The AI engineering stack we built internally – on the platform we ship
10 hours ago
- #AI Engineering
- #Cloudflare Platform
- #Developer Productivity
- Cloudflare achieved 93% R&D adoption of AI coding tools within 11 months, driven by a tiger team called iMARS.
- Key metrics include 3,683 internal users, 47.95 million AI requests, and 295 teams using agentic AI tools over 30 days.
- The architecture integrates Cloudflare products like AI Gateway for routing, Workers AI for inference, and Access for authentication, all accessible via a proxy Worker.
- AI Gateway centralizes LLM requests, providing cost tracking and Zero Data Retention, with 20.18 million monthly requests and 241.37 billion tokens routed.
- Workers AI processes 51.83 billion tokens, offering cost savings and low-latency inference, especially for tasks like documentation review and security agents.
- MCP Server Portal with Code Mode reduces token overhead by collapsing tools, improving efficiency as the number of servers grows.
- Backstage serves as a knowledge graph with 16,000+ entities, enabling agents to access service ownership, dependencies, and documentation.
- AGENTS.md files generated for thousands of repos provide context on codebase structure and conventions, improving AI accuracy.
- AI Code Reviewer automates merge request reviews, categorizing findings by severity and citing Engineering Codex standards.
- Engineering Codex standardizes rules, available as an agent skill for compliance checks and audits.
- Future plans include background agents using Durable Objects and Sandbox SDK for cloud-based development tasks.