• 400 Gb/s optical transmitter - simulation and wavelength control

    Building a physics-based simulator of a 400 Gb/s optical transmitter with eight ring resonators. Using it to develop controllers that maintain wavelength lock despite thermal crosstalk and laser drift.

    Hand-drawn ring resonator model showing two couplers, clockwise and counterclockwise fields, backscattering, and four ports. Eight-ring transmitter simulation showing monitor photodiodes, drivers, and a thermal disturbance beside the first ring. CAD view of eight ring resonators arranged along a shared optical waveguide.
  • Maskless lithography stepper - 3D CAD, optical simulation & mechanical design

    Designing a complete maskless lithography stepper targeting <1 µm features, with 10× projection optics and a micro-manipulator stage.

    Full CAD assembly of the lithography stepper, showing the projector, microscope objective, gold micro-manipulator stage, and aluminum frame. Close-up of the two projected mirrors on the chip, with coloured rays converging at their corners. View 1 of two mirrors, showing their light paths through the objective to the chip and camera. Fourteen CAD parts for the stepper, arranged in seven columns and two rows.
  • Diffusion models from scratch

    Wanted to understand how diffusion worked (also the inspiration for my website background), so I learned all the theory and coded up a progression from flow matching to diffusion, then CFG and EMA.

    Ten final color image samples generated by the diffusion model, extracted from saved notebook outputs. Four generated samples at six stages of sampling, progressing from random noise to recognizable images. Twenty handwritten digit samples generated by the diffusion model, extracted from saved notebook outputs. Complete saved notebook sampling grid showing ten generated images progressing from noise across all twenty-two stages, numbered zero through twenty-one.
  • Shady - GLSL to SPIR-V compiler

    Spent a lot of time working through the Dragon Book for a class. Ended up building a GLSL compiler that emits SPIR-V, then using a small Vulkan renderer to run the shaders on my GPU.

    Eight rows of sixteen teapots on one continuous dark workspace, with 64 simple and medium surface shaders followed by 64 complex shaders, compiled by Shady and rendered with Vulkan.
  • Reinforcement learning for microgrids

    Wanted to learn reinforcement learning and learned why it’s so hard to make a good reward function. Built a custom Gym environment, added dropout to Stable Baselines3, and wrote a paper.

    Microgrid system diagram showing solar PV, battery storage, generator, fuel cell, power converters, switchboard, controller, grid connection, and loads. Microgrid dashboard showing power production and consumption, solar generation, battery level, grid import, and weather forecasts. Original PPO notebook plot of solar, wind, generator, fuel cell, battery, and grid power during continuous operation. Original PPO notebook plot of critical, essential, and non-critical loads during continuous operation.
  • GPU-accelerated 3D map renderer

    Really wanted to learn CUDA back then, so I built a 3D renderer with frustum culling to achieve 240+ FPS.

    Mapcraft rendering of the Royal Ontario Museum and surrounding Toronto buildings, roads, and trees. CN Tower and Rogers Centre rendered from an elevated viewpoint over downtown Toronto. High aerial rendering of Queen's Park and the surrounding Toronto street grid. Close-up rendering of the Ontario Legislature, its green roofs, nearby roads, and trees.
  • Neural network on an FPGA

    Had a crazy idea to build a neural network in Verilog, basically writing the circuit for a model that classifies handwritten digits. This was one of my hardest projects. Learned the hard way that simulation != FPGA. Performs inference on MNIST in under 0.53 MB of total FPGA memory.

    Handwritten digit eight displayed on a monitor beside the DE1-SoC FPGA running the neural network. RTL schematic of the Verilog neural network. ModelSim simulation output for the neural network hardware.
  • Image captioning using vision transformers

    A friend challenged me to build an image captioner when I had no deep learning experience. Speedran the history of deep learning in my first year summer and built the entire thing by hand (back when LLMs were just as good as autocomplete).

    Tennis photograph with its original dataset caption above and the model's decoded caption below. Skateboarding photograph with its original dataset caption above and the model's decoded caption below. Surfing photograph with its original dataset caption above and the model's decoded caption below. Image captioning architecture showing a ResNet-50 image encoder, a custom transformer with image and word embeddings, and a dictionary tokenizer.
  • Dope interactive shell (dish) - a Unix shell

    Wrote a shell by hand in C during a small research gig in first year, for a course called ECE344 at the University of Toronto. Dove deep into xv6, got systems-pilled, and started developing a taste for operating systems.

    Dish running real commands, a 100,000-line pipeline, and a background process in a recorded terminal session.