Cooper Morgan

As a part of my graduate research, I'm replacing the traditional terrain rendering pipeline with neural networks. Instead of decoding massive elevation tiles at runtime, TINRS trains compact models (~0.5-1MB) that the GPU evaluates directly via Vulkan compute shaders. I went through about 6+ architectures before landing on multi-resolution feature grids: 30× compression over raw data, with accuracy within the source data's own measurement noise. The Zig viewer hits 20,000+ FPS on desktop and 300 FPS on a Raspberry Pi 5. Paper in progress.

My main graduate research: a modular framework for DNS exfiltration research. It includes an RL agent that learns encoding strategies, an LSTM timing model that mimics real traffic patterns, and Stegatetra, a steganographic encoding I designed that hit 0% detection across every detector we threw at it. Stegatetra encodes data through which queries get sent, rather than what's in them. Accepted at ACNS 2026.

A team project for our robotics course where we trained a DQN agent to autonomously seek a light source while avoiding obstacles on a Bitcraze Crazyflie 2.1+ nano drone. We trained in Webots simulation using Stable Baselines 3, then deployed the model on-device via ONNX on the drone's GAP8 RISC-V chip. The drone uses an AI-Deck camera for brightness-based bearing detection and a MultiRanger deck with 5 ToF sensors for obstacle avoidance, all within a 29g takeoff weight budget.

The image is from an early training run where the drone kept flying straight at a person when the light source was out a glass wall.

This project is a linter and syntax highlighter for KPL (Kernel Programming Language), designed to enhance the experience for students who are forced to use this... language. After being disappointed with the syntax highlighting and little to no support for KPL in VSCode, I decided to create my own extension. I also spent way too much time on the linter and syntax highlighter for the "blitz assembly" language that KPL compiles to.