I write code that sits between machine learning and systems work. Lately that has meant local vision models driving hardware, GPU kernels, and cryptography.
A cipher where the ciphertext carries a mathematical formula's identity and its parameters instead of plain text. There are three layers. The engine is deterministic and does the encryption. A small neural network sits beside it as an auditor, trying to tell one ciphertext from another, and it never touches the cipher path. The third layer answers questions about the formula corpus.
The repository also holds a constant-time X25519 core written in C, a fuzzer
that follows branch coverage through sys.monitoring, a cycle-level timing
harness, and a tool that reads the compiler's assembly output to check the
constant-time code survived optimisation. There are 29 decision records
explaining the reasoning, and the audit document says plainly which claims were
measured and which were not.
GPU-Accelerated-Template-Matching
Normalised cross-correlation template matching, written as PyTorch GPU kernels.
Vision-Guided-4-Axis-Robot-Arm-with-Object-Detection
A four-axis arm that finds an object, tracks it and picks it up. A local Qwen vision model does the seeing, an Arduino Uno drives the servos.
An eye that follows what the camera sees, again with a local Qwen model and an Arduino behind it.
A screener that ranks US small and mid cap equities against 28 factors, run daily.
Writing code whose timing does not depend on secrets, and then measuring whether that is actually true rather than assuming it. Fuzzing. Keeping statistical models away from the parts of a system that have to be exactly right.