Deep dives into engineering, AI, and modern infrastructure. We document our learnings and architectural decisions. No fluff, just high-signal content.
GitHub published three Moderate advisories for the MCP Python SDK on October 27, 2026. Two are fixed by upgrading to 1.30.0 or 2.2.0. The token audience check stays off until you turn it on, and the redirect fix can break clients that work today.
TypeSafe's Jev returns typed decisions with probabilities instead of text, at 70 to 500ms and $0.042 per million input tokens. Here is what it replaces, what it does not, and how to test it before you trust it.
A practical guide to retrieval-augmented generation with hybrid search, cross-encoder reranking, and citation grounding.
How we achieve atomic deployments with zero cold starts using Workers static assets and gradual rollout strategies.
What breaks when Model Context Protocol servers go to production: authentication, versioning, timeouts, observability and security.
Why we default to the edge for new projects, and the specific scenarios where you still want a traditional origin server.
When does fine-tuning beat prompting? We break down training and serving costs, latency tradeoffs, and quality deltas, using published GPU prices and worked estimates.