MCP in Production: Designing Tools an AI Can Use Safely
An open protocol can make your integrations reusable. The interesting work starts when you decide what those integrations are allowed to do.
Writing · Nadeera Gayantha
Engineering notes,written after the demo.
Long-form explanations of the decisions that make software hold up in production: how AI systems reach real data, how retrieval finds the right evidence, and how APIs keep their promises.
- Articles
- 3
- Reading
- 35 min
- Updated
- 2 Oct 2026
Three-article series
Building AI systems that hold up in production.
MCP, agentic retrieval, and schema-first APIs. Practical explanations for the decisions that matter after the demo.
- 01MCP in Production: Designing Tools an AI Can Use SafelyDesign useful tools and enforce clear limits.10 min
- 02Beyond Vector Search: Retrieval That Respects Business RulesFind evidence that satisfies the whole question.13 min
- 03From as MyType to Real API ContractsValidate data, authorize actions, protect state.12 min
Beyond Vector Search: Retrieval That Respects Business Rules
Hybrid search, rank fusion, and reranking can improve your evidence. Production retrieval also needs to respect the parts of a question that aren’t negotiable.
From as MyType to Real API Contracts
How to combine TypeScript, runtime schemas, authorization, and database guarantees without pretending one of them can do every job.
New notes are published here first.