Co-authored technical article demonstrating a system for detecting environmental changes across multiple drone flights using monocular VGGT-SLAM and semantic segmentation. The approach reconstructs 3D scenes from single-camera video, identifies objects with ML, and aligns separate reconstructions to highlight what persists, appears, or disappears between flights.
A daily-use Obsidian vault template centred on Templater for note creation and Bases views surfaced in the sidebar. Keeps the plugin footprint small, with Excalidraw as the main optional addition.
Project I supervised exploring a lightweight, offline voice command system for a Mars rover. The system converts natural speech into executable commands by chaining hotword detection, speech-to-text, and a local LLM — enabling hands-free rover control without network access or a traditional gamepad. Written by Pierre Hellich.
Transformer-based classifiers for detecting hateful and offensive content in short texts. Tackled cross-lingual classification using joint-learning and data translation approaches, outperforming existing multilingual models on zero-shot benchmarks.
Evaluated state-of-the-art techniques for audio denoising and super-resolution, then applied generative adversarial network methods from image processing to the audio domain.
Data story exploring potential bias in Amazon user reviews and methods to correct it, working on 20 GB of comments using Pandas, PySpark, and Matplotlib.
3D maze game built in C with OpenGL, featuring world physics, lighting effects, rendering optimizations, a user interface, and sound effects.
Multi-class perceptron POS tagger (93%+ accuracy) and sentence dependency parser (76%+ accuracy) implemented from scratch in Python.
Android app for mental arithmetic practice with multiple modes including test mode, timed mode, and rush mode.
Multiplayer iOS implementation of Dots and Boxes supporting up to 4 local players, with an AI opponent. Written in Objective-C.