Writing
Agentic systems in production.
Technical writing on system design, retrieval architecture, orchestration patterns, and the decisions that separate production AI from prototypes.
Why your RAG pipeline works in staging and fails your users
You tuned retrieval for two weeks. The demo impressed the team. You shipped it. Three weeks later, users are getting answers that have nothing to do with their questions. Here is what actually happened.
The orchestration layer that doesn't collapse under its own weight
It starts clean. A planner, three tools, a simple loop. Six months later you can't explain what it does without opening four files simultaneously. This is the pattern that prevents that.
Memory in agentic systems: four layers most teams collapse into one
Your agent forgets things it should remember. Remembers things it should forget. Starts every session cold. The fix is not a bigger context window — it is designing memory intentionally across four distinct layers.