Took on a Kaggle competition for agricultural weed detection and extended it significantly beyond baseline. Applied transfer learning with modern vision architectures and ensemble techniques. Explored parameter-efficient fine-tuning methods to improve model performance.

Investigated how emotion and sentiment influence engagement on Twitter during COVID-19. Extracted polarity, VAD, and 11 emotion categories from tweets using both lexicon-based methods and transformer models. Decomposed text embeddings into orthogonal affect/non-affect subspaces to isolate emotional content contribution.

Built for fun—a constraint satisfaction solver with a polished command-line interface. Clean, navigable TUI for an elegant terminal experience. Currently cleaning up for GitHub release.

Custom hardware build: OAK-D Lite camera on motorized turntable. ESP32 MCU controlling stepper motor with rotary encoder interface. Parameterized capture sequences for different scanning modes.

Complete machine build: aluminum frame, steel components, custom electronics. Configured grblHAL firmware with TMC2209 stepper drivers. Raspberry Pi web interface for remote operation and job management.