Writing
Notes on data, agents, and trust.
- ·10 min read
You logged everything and still can't prove who did it
Signing solved integrity. Article 12 also demands traceability, and across systems that becomes an entity resolution problem almost nobody in the governance conversation is building for.
- ·11 min read·onlymetrix.com
We ran the same benchmark as dbt Labs. Here's what they didn't measure.
Same benchmark, different question. What a SQL generation score hides once you measure governed metrics instead.
- ·9 min read·onlymetrix.com
Deterministic vs probabilistic data access
Two architectures for letting an agent reach a warehouse. Who writes the SQL decides everything downstream of it.
- ·9 min read·onlymetrix.com
Harness engineering for data access
Constraining what an agent can do beats prompting it to behave. The harness is the control surface, not the model.
- ·10 min read·onlymetrix.com
Why AI agents shouldn't write SQL
LLM-generated SQL fails on reliability, security, and governance at once. Context injection narrows the gap, it does not close it.
- ·8 min read·onlymetrix.com
Your metrics haven't changed. Why is your AI re-deriving them?
AOV and LTV were settled years ago. Recomputing them from scratch on every query is a governance failure dressed up as a feature.
- ·10 min read·kanoniv.com
Declarative identity resolution vs traditional MDM
Why entity resolution belongs in version control next to your models, rather than inside a monolithic platform.
- ·11 min read·kanoniv.com
How we built a Rust-based identity resolution engine
The architecture behind Kanoniv's matching core, and where the performance actually came from.
- ·14 min read·kanoniv.com
Why most multi-agent systems fail in production
Identity fragmentation, not model quality. When agents cannot agree on who a record refers to, memory and coordination break down.