Writing
What If Learning Software Engineering Felt Like Joining a Real Team?
Watching someone build software is not the same as learning how to work inside software. I started building noetherkin to make learning engineering feel more like joining a real team. Updated for noetherkin's current CLI and skills architecture.
NoetherKin: Architecture Notes, Part 2 — Skills, Catalog, and What's Actually Built
Part 2 of 2. The 21 portable Agent Skills, the 34-track/87-project/7-forge catalog, a worked task from assignment to evidence with real CLI commands, the safety boundaries the CLI enforces, and an honest accounting of what's shipped.
NoetherKin: Architecture Notes, Part 1 — Protocol, CLI, and Workspace
Part 1 of 2. NoetherKin's structure changed a lot since the first draft of this piece: it's now a released CLI, a trusted single writer, and twelve closed schemas, not a folder of intentions. Workspace layout, install flow, and the protocol's state model.
What I Learned Profiling PyTorch Memory Leaks Across Two Backends
Eight leak patterns on MPS and CUDA — CUDA dampens leak signals but does not hide them, and the bugs that matter show up everywhere.
Diagnosing a PyTorch Memory Leak on Apple MPS
How Stormlog helped us find and fix a subtle GPU memory leak that only appeared on Apple Silicon.
I Helped Build a GPU Memory Profiler. Then I Had to Learn What GPU Memory Actually Is.
Walking through the Stormlog tutorial after shipping the tool — from PyTorch counters to a deliberate leak, OOM evidence, and the fix.
OOM Flight Recorder Ring-Buffer: Deep Dive
Why a ring-buffer flight recorder for GPU OOM fills a gap PyTorch snapshots and NCCL tracing do not — temporal context, automatic dumps, and structured artifacts.