Project Portfolio

This independent research project explored enabling AI on resource-constrained environments (smart devices, smart phones, etc.). The effort highlights knowledge distillation, a powerful technique used to compress neural networks into smaller, faster, and more efficient versions

This independent study built a simple LLM app that wraps a hosted model, programmatically interacts with the LLM context and conversation history, and generates LLM responses with proprietary context.

Built a facial-emotion classifier using transfer learning and fine-tuning to predict seven emotions from grayscale images. The model accuracy improved over baseline and surfaced real-world challenges and considerations: how well the transfer model’s learned features matched the new data, and the quality of the new image data to be classified (class balance, low resolution, and ambiguous expressions).

Provide a repeatable, low-friction workflow to run NVIDIA GPU-accelerated TensorFlow workloads on Windows 11 using Ubuntu & WSL 2.

Enhanced an existing open-source vector search application (MycoSearch) to enable rapid mycological species identification via search interface. Generated a custom class ("Taxon") to format species content from annotated text and embedded into search corpus.

Building on SKOL Part 1, this effort automated the parsing of unannotated research literature, classification of paragraph content using machine learning techniques, and conversion of classified paragraphs to the custom "Taxon" format for embedding in the MycoSearch application (SKOL Part 1) .

Predicted student outcomes (dropout, enrolled, graduate) from demographic and academic history data using machine learning techniques. Surfaced early-warning patterns and actionable predictors universities can use to support at-risk students.

Analyzed the solar setup of a single-family home in Hawaii to quantify capacity, grid reliance, and ROI, with special focus on EV charging. The analysis showed the system can support household loads, but EV charging drives reliance on the grid. Recommendations were surfaced for adjusting charging schedules and scaling the solar system capacity over time.

The Name Navigator database was designed to provide a culturally inclusive name personalization system that could integrate into Customer Relationship Management software to enable a nuanced handling of multilingual and multi-format names. This project focused on demonstrating how thoughtful database structure can improve inclusion, reduce bias, and enhance client trust.