A few things I've built, from bare-metal systems to genomics to teaching.
a processor, from the ISA up
A RISC-V RV32I processor core in RTL: fetch, decode, register file, ALU, the rest. It compiles ordinary C programs and runs them under Verilator. I built it for Stanford's advanced systems course, and I'm now refining the design with Prof. Mark Horowitz, with the goal of an actual silicon tape-out.
Built without AI
CYW43438 on bare metal
Bringing up the Broadcom/Cypress CYW43438 on a Raspberry Pi with no OS: SDIO, Cortex-M3 firmware upload, and a small network stack (WPA2, DHCP, UDP), reverse-engineered from Linux kernel source and datasheets. Mostly this meant reading the datasheet until the hardware finally answered, then documenting the whole bring-up.
Used AI to reverse-engineer Linux kernel source
a bare-metal game with a physical controller
A Subway Surfers-style game written in C on a Mango Pi (RISC-V), bare metal with no OS or libraries. You switch lanes by tilting a real skateboard wired to an MPU-6050 gyroscope, read over I2C with interrupts; obstacles spawn on a timer and the graphics run on a model-view-controller setup. Built with Finley Desai.
Built without AI
single-cell leukemia sequencing pipeline
A ground-up C++17 rewrite of a high-throughput pipeline that turns raw FASTQ reads into per-cell genotypes and antibody-marker profiles, with quality-control metrics and visual summaries. Mission Bio's Tapestri software is a black box, so I reverse-engineered how its raw reads encode each cell (the DNA-amplicon and antibody-tag barcoding) and built the pipeline to demultiplex the data itself. It runs on Slurm-based HPC, parallelized across samples and targets for about a 10× speedup, which lets researchers follow a patient's samples from diagnosis through remission to relapse. Built at UCSF.
I built most of it; AI used to extend it
a tutoring curriculum
A single-variable calculus curriculum I built for my students, taught alongside physics so the intuition comes before the machinery. The throughline: start from where an idea comes from, not the formula it ends up as.
I built most of it; AI used to extend it
a framed information display
A wall-mounted Raspberry Pi dashboard on a Waveshare e-ink panel: weather, calendar, and the day at a glance, rendered as a crisp 1-bit image with Pillow and driven over SPI. Mostly an exercise in doing a lot with 648×480 black-and-white pixels.
Built with AI
a self-tracking dashboard
A local dashboard that pulls my Oura and Hevy data through their APIs into one view of sleep, readiness, and training, with a small Flask backend and a React board for course work that auto-imports assignments from GitHub. A study in gluing a few APIs together cleanly: live fetches, a short cache, atomic JSON writes.
Built with AI
founder
An earlier venture in high-end baked goods: I built the brand, recipes, and packaging end to end, and negotiated a discounted shipping contract with UPS. Where I learned that shipping a product means shipping everything around it too.