Jane Street Applied ML Puzzle
"Puzzle"
Jane Street Applied ML Puzzle
"Dropped a Neural Net"
Kyle Lukaszek
Summary
Hi, I'm Kyle. I like machine learning, computer graphics, programming language theory, and solving puzzles.
I'm finishing an M.Sc. in Computer Science at the University of Guelph, where my thesis focuses on efficient, interpretable self-supervised models of human visual perception and what we can learn from our eyes.
Outside of my research, I've built Kiln, a Rust-based Vulkan 1.3 and Metal 4 render hardware interface that powers my toy spectral path tracer Spectra. I have also worked on developing Haskell bindings for the SDL3 library, and an analytically constructed transformer that drives the CPU and RAM of a playable NES emulator.
Languages
GPU APIs
Libraries
Tools
hover / tap for a surprise
Achievements
Publications
GI '26: Graphics Interface, Waterloo, ON, Canada
GI '24: Graphics Interface, Halifax, NS, Canada
Projects
An analytically constructed transformer that runs a complete MOS 6502 inside a console-accurate NES emulator.
- Compiled the complete 6502 instruction set into transformer weights without training. The residual stream holds CPU state, while an attention-addressed write log provides memory.
- Built a native C++ NES emulator around the transformer CPU, including the PPU, APU, system bus, DMA, controllers, cartridge loading, and NROM and MMC-family mappers.
A cross-platform Rust render hardware interface for modern graphics and compute workloads.
- Designed a single Rust API for acceleration structures and compute, mesh, vertex, and pixel shader pipelines across Vulkan and Metal.
Neumours
A compact glioma segmentation model paired with a differentiable real-time 3D viewer.
- On BraTS 2023+, the model reached 0.823 Whole Tumour Dice using 1/118th as many parameters as nnU-Net. It processed each case in about 2.8 seconds, compared with 54 seconds for nnU-Net.
- Built a realtime MRI volume renderer that caches predictions on the GPU.
Complete Haskell FFI bindings for SDL3's native application and GPU APIs.
- Contains 59 examples, including GPU examples that run across SDL's Vulkan, Direct3D 12, and Metal backends.
A WASM32 build of Google's Tint shader compiler and SPIRV-Tools.
- Ported the WGSL and SPIR-V portions of Tint and SPIRV-Tools to WASM32 for browser use.
- Exposed JavaScript and C++ APIs for browser-side shader cross-compilation, with the same runtime available as a static library.
Browser-native GPU applications for large LiDAR datasets, colour science, and other things.
Experience
Teaching Assistant
CIS*2750, CIS*3090, CIS*3110 · University of Guelph
Jan 2025 – Apr 2026
- CIS*3090 Parallel Programming F26: Returning to help teach parallel programming.
- CIS*3110 Operating Systems W26: Taught OS concepts, system calls, Bash commands, and shell scripting weekly.
- CIS*3090 Parallel Programming F25: Taught Pthreads, OpenMP, OpenCL, C CUDA, and Nvidia Warp.
- CIS*2750 Software Systems Development W25: Taught C, TUI development, CPython FFI integration, and SQL.
Research Assistant
HCI: Hardware & Modelling · University of Guelph
Jan 2024 – Jan 2025
- Built a cross-platform JETI Spectraval 1501 driver replacing the official solution, reducing spectrometer latency by more than 50%.
- Co-authored peer-reviewed work accepted at Graphics Interface 2024.
Research Assistant
NLP: Pre-training & Post-training · University of Guelph
May 2023 – Sep 2023
- Accelerated preprocessing pipelines, trained, and fine-tuned a variety of models on a large Twitter corpus.
- Worked with C CUDA, PyTorch, and the SciPy ecosystem.
Education
M.Sc. Computer Science
University of Guelph
Supervised by Dr. Denis Nikitenko and Dr. David Flatla. Research focus on perceptual colour science and inverse problems.
Coursework
Honours Bachelor of Computing
University of Guelph
Area of Application in Mathematics. Dean's List 2023 & 2024. Research Assistant under Dr. Denis Nikitenko and Dr. David Flatla.