At TEXMiN, I worked with the R&D team on classifying 76 minerals from 256-channel hyperspectral imagery for mineral deposit identification. I designed a workflow combining data augmentation, PCA/UMAP visualization, and 2D/3D CNN and Vision Transformer models.
At Brane Enterprise, I contributed Python/Django backend components for the Stringer chatbot in Enterprise Control Tower (ECT).
On Adaptive MambaVision – Input-Adaptive Feature Fusion, I extended MambaVision with token-dependent gates across six mixers. On an additional 5,000-image CIFAR-100 evaluation sample, I observed 82.84% accuracy versus 81.88%, while documenting mixed gains and computational overhead.
On CleanQuad – Autonomous Quadruped Litter Collection, I developed simulated Unitree A1 locomotion and fine-tuned YOLOv8n for litter detection. The perception model achieved 0.965 mAP@0.5 versus 0.342 for a color-based baseline on 500 unseen-layout test images; the integrated team-built simulation collected 7/7 reachable litter items.

