Containing a Frozen Framework Until You Can Leave It
August 18, 2026
Rebuildability, a watched dependency tree, tight isolation, and an exit plan with dates. How to run unsupported software on purpose instead of by neglect.
Read moreTutorials
Field notes from migrating frozen ML systems.
PyTorch or ONNX: Choosing a Target for an MXNet Port
July 21, 2026
PyTorch is a home; ONNX is a shipping container. The choice turns on one question — will this model keep learning? — and the escape hatches are not symmetric.
Read moreBuilding a Numerical-Parity Harness for a Framework Port
June 23, 2026
Not "does it work" but "is it the same model". The input corpus, the comparison contract, the execution protocol, and how to localize divergence when parity fails.
Read moreWeight Conversion Pitfalls: MXNet to PyTorch
May 19, 2026
The dangerous conversions are the ones that almost work. BatchNorm semantics, padding, layout, custom operators, and preprocessing — the failures that survive a smoke test.
Read moreMXNet Is in the Apache Attic: What That Means for Production
April 14, 2026
Attic status is not a pause. No releases, no security patches, no community — here is what that translates to operationally, and the sequence that gets you out.
Read more