← All case files
Prompt logs as a data-warehouse problem
An analytics platform for LLM prompt telemetry: about 4.6 million raw execution events normalised, feature-engineered, warehoused and mined.
- Raw logsunstructured
- ETLadapters and validation
- Feature engine15+ attributes
- Star schemaviews and indexes
- Analytics APIand a dashboard
- Teams tune prompts by feel. The logs that could show what actually works are unstructured and huge.
- Millions of events, one machine, and queries that have to stay fast.
Decisions.
Treat prompt engineering as observability.
Logs go through schema adapters and validation before anything is counted, so the analysis sits on data you can trust.
A warehouse, not a pile of files.
A PostgreSQL star schema with materialized views and indexes supports OLAP-style questions in real time.
Start with boring models.
K-Means clustering, Apriori association rules and a baseline logistic regression tell you whether there is signal before you reach for anything heavier.
- A FastAPI layer over the warehouse and a dashboard on top, so the backend data work is the product and the UI is a view of it.