I contribute to open source as goingforstudying-ctrl, working on machine learning and AI engineering. That includes ML framework correctness, LLM inference, retrieval pipelines, and MCP tools. I also spend time in the distributed backends these systems rely on.
I’m drawn to bugs where the original code makes a convincing argument for itself. TensorFlow reused a forward operation to calculate a gradient. Swiftide cached a document so it wouldn’t have to process it again. Following those decisions far enough explains both their appeal and their failure.
These notes are where I put that longer explanation. I want a reader to be able to follow the failure, understand the change, and judge whether the regression test earns its confidence. Each technical post links to the merged public patch.