TabICL — recategorizing 694 books with a calibration gate
A tabular foundation model beats a tuned baseline by +0.10 f1-macro with 4.4× better ECE. The gate evaluates both models on identical folds and reasons on the paired delta.
Read →Case studies
Context, constraint, decision, measured result.
The format is always the same: the context, the constraint that ruled out the obvious solution, the decision taken, what pushed back, and the measured result. Without the part about what pushed back, a case study is just an advert.
Supply chain: 122 → 0 CVEs
Four production environments, 122 known vulnerabilities, and a real root cause nobody would have found by reading application code. The literal demonstration of the Audit engagement.
Read →Pooling: 208 GB of redundant RAM removed
Nine applications each loading their own models on a 128 GB machine. Why the answer was not to write one more application.
Read →Your data cannot leave the building?
That is precisely the problem I solve. A 30-minute call is enough to scope an audit.