What we're learning while building the intelligence layer — architecture decisions, failure modes, and patterns that actually work in production.
How we extract business logic from query history, dbt lineage, and pinned Slack threads without ever ingesting customer records.
Filtering after generation is too late. We explain why scope enforcement has to happen before a single token is produced.
Intent-aware composition beats context stuffing. Here's the benchmark methodology and where the savings actually come from.
A walkthrough of how conflicting metric definitions compound across an org — and what resolving them changed downstream.
Lessons from building an MCP server that agents query thousands of times per hour, including the caching model we settled on.
Version control, pull requests, and CI for business definitions. Why GitOps applies to your semantic layer too.
Roughly twice a month. Engineering notes, no marketing filler.
