AI Integration
The interesting problems in AI integration are almost never the prompt. They are retrieval quality, what happens when the model is wrong, and whether the unit economics survive contact with real usage.
How I use it
Retrieval that returns plausible, not correct
Vector similarity alone returns neighbours, not answers. Useful retrieval means deliberate chunking and combining embeddings with structured filters so the model is reasoning over the right rows.
Cost that scales past the point of sense
Routing each task to the cheapest model that clears the bar for it. Most pipelines have a long tail of calls that never needed the frontier model and a short head that genuinely does.
Output reaching a customer unchecked
Evaluation gates before anything user-facing ships. Without one you are not running an AI feature, you are publishing whatever the model produced that day.
Multi-tenant data with a model in the middle
On Brand KB the tenant boundary is enforced in the query itself, so isolation does not depend on the prompt behaving.
What I use
- Claude API
- pgvector
- Model Context Protocol
- Drizzle ORM
Related skills
Want the detail behind any of this?
I can walk through any of it in as much depth as you want.