01
The test was protecting the bug
TensorFlow returned a zero gradient, and an existing test agreed. A straight line through the origin gave me a reason to doubt them both.
4 min read
MLE & AI Engineering
The test was
protecting
the bug.
TensorFlow returned the wrong gradient.
An existing test expected it.
I work on ML frameworks and the systems around them. These are the bugs I followed, the patches that merged, and the reasoning a diff can't quite hold.
Read the TensorFlow storyFig. 01 / TensorFlowAt the origin
Forward value0.0000
Correct gradient1.1314
Old gradient0.0000
A small counterexample.
A surprisingly persistent assumption.
01 / From the debugging notebook
Follow the failure.
Four stories, with the code
and tests behind each one.
02
The retry skipped the document that failed
An embedding failure should leave a document available for retry. Swiftide's cache had already decided it was done.
5 min read
03
The socket race started in a destructor
I followed a Redis socket race out of Celery's polling loop and into result cleanup. The fix needed a lock that could survive callbacks and a fork.
4 min read
04
A late success sent Airflow down the failure path
The worker's success report was accurate. By the time the scheduler read it, the task had already resumed, and the state guard was one state short.
4 min read
02 / Beyond the write-ups
A few merged patches.
47merged public PRs
Across 28 repositories.
ML frameworks, AI infrastructure,
and distributed backends.
Count checked Oct 11, 2026.