At HAAG’s NAHPU Project, I’m benchmarking lightweight, on-device deep learning models for insect specimen coloration studies and evaluating embedding clustering. I’ve also evaluated segmentation models to isolate specimen regions from noisy museum photography.
In my computer vision loss-function research, I trained a FixMatch pipeline on CIFAR-10 and compared Poly-1, LACE, and Proportion Loss across label ratios. My results showed Poly-1 generalizes to balanced and imbalanced data distributions, and I drafted the loss-functions section of the research paper.
I fine-tuned TinyLlama-1.1B on medical flashcard Q&A pairs with LoRA, reducing trainable parameters by approximately 99% while preserving task performance. I also built a retrieval-augmented Q&A pipeline with PubMedBERT embeddings and a persistent FAISS vector store, so it can retrieve medical documents and generate responses without external APIs.
For a separate CIFAR-10 project, I designed an 11-layer CNN with residual connections and an Inception module. I also implemented 2D convolution and manual backpropagation in NumPy.
