Short pieces on agents, applied ML, evaluation, and the production constraints that shape our research.
One agent asked to plan, build, test, and validate still fails on large work the way large human projects fail. Skailr is the org structure we install on top of Claude Code and Cursor.
Read the note§ 02 — Archive
Matching pipelines, scoring models, and agentic systems all fail on the same class of case: the messy one a tidy demo never shows you.
Read the noteAd-hoc scraping gives you text about the market. An ontology gives you a structure you can query, reason over, and eventually act on.
Read the noteRouting, planning, and adapting across agents is only trustworthy if you can see it happen. Autonomy without observability is a black box that occasionally works.
Read the noteTest data is only useful where your pipeline can read it. Export formats and a programmatic API matter as much as the data itself.
Read the noteEntity resolution systems are scored on the duplicates they catch and the false merges they avoid. A test set of unique, well-formed rows cannot exercise either.
Read the noteGeneric fake data looks fine in demos and breaks when you test record matching systems. Here is what entity resolution evaluation harnesses actually need from synthetic test data.
Read the note
