Director of Machine Learning at Cloudastructure. Computer vision for physical security; GPU performance and the unglamorous parts of ML systems.
On the side I build and ship my own tools: Unfocus, Swage, and Pythonlings.
I write about CUDA, NVIDIA driver internals, and numerical stability at abhik.ai.
- Vision Transformer Deep Dive: ViT to Modern Descendants Aug 18, 2026
- Xid 31 MMU Faults: What Causes Them and How to Fix Production GPU Crashes Jun 01, 2026
- The Complete NVIDIA Xid Error Field Guide Apr 21, 2026
- CUDA Matrix Multiplication Optimization: From Naive to Near-cuBLAS Apr 07, 2026
- Numerical Sensitivity: Why FP16 Breaks NAdam Jan 08, 2026
- Unfocus: a break reminder that asks one thing of you, look at something far away. Installable via Homebrew and apt.
- Swage: a Python-embedded MLIR/LLVM GPU compiler that turns variable-sized dense segments into efficient GPU tile tasks.
- Pythonlings: Rustlings-style Python exercises in a live terminal TUI, with practice tasks and offline Python docs.
- Researching failure modes in ML training clusters (GPU, network, and storage); wrote the NVIDIA Xid Error Field Guide from that work.
- EuroPython 2026: "An Introduction to Writing Fast GPU Code in Python".
- PyCon Italia 2026: "Write Your First High-Performance GPU Kernel in Python!" (github).
- PyCon India 2025: ArrPy: rebuilding NumPy from scratch.





