I'm Jake Kasper — I build low-level software: kernels, allocators, schedulers, and the unglamorous plumbing that makes hardware do useful things.
Most of my work starts with a question like "but how does that actually work?" — which is how I ended up writing an operating system for fun.
I care about correctness, small dependency footprints, and code that a future maintainer can actually read. I'm happiest somewhere between the hardware and the abstraction.
Outside of kernels, I'm interested in compilers, storage engines, and developer tooling.
Machine learning is part of the story too — I co-authored a peer-reviewed paper on deep-learning methodology for EEG as a freshman, and I still build neural nets from scratch to understand them end to end.
A hobby kernel written from scratch in C and AArch64 assembly. Built to drive Raspberry Pi hardware.
Boots on a physical pi4b. Implements a 4-level virtual memory system, a buddy page allocator and slab-based kmalloc(), a preemptive scheduler, copy-on-write based process replication, an ext4 driver over EMMC, an ELF loader that runs real user binaries within the ext4 filesystem, and a UNIX-based syscall table for compatibility with existing software, such as python and GNU.
Features a simple command line interface for user interaction, either via serial/UART for input/output and touchscreen over DSI for external output (supporting a subset of ANSI sequences).
Debugging enabled over JTAG with J-Link + OpenOCD + GDB against a running board — the same trace scrolling behind this page.
screen. Servo on the Uno's D9.Machine learning, data science, and systems since 2021 — internships, research, and a full-time ML engineering role.
A custom RISC-V bootloader brings up an empty board, then runs two coordinated harts: a scheduler that builds offline Rate-Monotonic and Earliest-Deadline-First schedules, and a task manager — talking over a hand-defined inter-hart messaging zone.
A feed-forward net with zero dependencies: a hand-written matrix library, ReLU / sigmoid / softmax activations, MSE and cross-entropy losses, and full backpropagation — including softmax Jacobians for batched training.
Machine-learning methodology, from my time in neuroscience research.
G. Brookshire, J. Kasper, N. Blauch, et al. · Frontiers in Neuroscience · 2024
A methodological study of how data leakage inflates reported accuracy in deep-learning models of EEG — and the safeguards needed for results that actually generalize. Co-authored as a college freshman.
Whether it's a role, a collaboration, or you just want to talk about page tables — my inbox is open.