Current

• DinoDreamer - Feature-visualization instrument for frozen vision models. Gradient-ascent synthesis inside DINOv2 — archetypal-SAE targets, MACO/Fourier/pyramid parameterizations, and measured render-and-reencode evidence for every run. (2026–present)
• SubspaceLens - Training-free interpretability for diffusion transformers using MCR² coding rate metrics. Measures subspace geometry — no SAEs, no learned components, just eigenvalues. (2026)
• Latent Dance - Real-time audio-reactive SDXL with SAE steering. 20,480 interpretable features driven by music at 50 FPS on a single consumer GPU. Music visualization as interpretability research. (2025–present)

Past

• SAM_CAM_BAM - Interactive music visualizer combining computer vision and audio processing for dynamic visual effects using webcam input and real-time audio analysis. (2025)
• hambaJubaTuba - Transforms audio features into latent space trajectories, producing temporally coherent music visualizations via diffusion models. The predecessor to Latent Dance. (2024)
• Better Spotify - Fullstack Rust web app that provides true random shuffling for Spotify playlists. (2025)
• workTimerPet - A little work timer pet that records work time and break time. (2025)
• A Careful Look into Graph Contrastive Learning - Experimentation testing claims on Graph Contrastive Learning and exploring improvements. (2022)

Publications

• "Exploring Cost-effective Computer Vision Solutions for Smart Transportation Systems" - Connected Cities for Smart Mobility toward Accessible and Resilient Transportation Center (2023)
• "A New Curb Lane Monitoring and Illegal Parking Impact Estimation Approach" - Transportation Research Part A (2022)