Explainable AI Research: ConvKAN vs. CNN Performance
Winter 2026 · Lead Researcher · Deep Learning & Mathematical Modeling
PyTorchConvKANHealthcare AIXAI
Healthcare models often need to work with limited labeled data, and black-box CNNs make it hard to trust what the network learned. This project compares Convolutional Kolmogorov-Arnold Networks (ConvKAN) with standard CNNs on malaria blood-smear classification across multiple dataset sizes, asking whether KANs can generalize with less data while remaining more interpretable.
CNN vs ConvKAN on malaria classification: parameters, accuracy, F1, training time, and learning curves
Built and trained CNN and ConvKAN baselines in PyTorch on malaria blood smear samples at different dataset sizes.
Modeled non-linear patterns with learnable activation functions on network edges, avoiding the black-box problem of standard MLPs.
Evaluated whether ConvKAN maintains accuracy under low-data regimes compared with conventional convolutional models.
Used the comparison as a foundation for more transparent, explainable tooling in medical imaging workflows.
Generative AI for Indigenous Language Preservation (Dakelh)
Fall 2025 · Lead Developer · Generative AI & NLP
NanoGPTHugging FaceBPELow-resource NLP
Dakelh is extremely data-scarce, so standard LLM tooling does not transfer cleanly. I built a small generative language model on a curated corpus (including the Dakelh New Testament translation) with a custom BPE tokenizer suited to the language’s morphology. The model can reproduce distinctive Dakelh surface patterns and form words from tokens. It cannot do complex reasoning or reliable conversational understanding.
Sample generations in a simple chat UI: pattern replication, not complex thought
Built a training pipeline with NanoGPT and Hugging Face Transformers under severe data constraints.
Authored a custom Byte Pair Encoding tokenizer so the model can assemble Dakelh-like word forms from subword tokens.
Demonstrated that the model can mimic unique Dakelh linguistic patterns in generated text, while remaining a limited pattern model rather than a general-purpose assistant.
Gamified ASL Fingerspelling System
Fall 2025 · Social Robotics Capstone
MediaPipeDjangoComputer VisionCustom gesture model
Off-the-shelf gesture models struggle with fine-grained ASL alphabet recognition, especially dynamic letters like J and Z. I designed a MediaPipe + Django pipeline that gamifies fingerspelling practice, trained a custom recognizer for all 26 letters, and extended it with temporal landmark sequences for motion gestures.
Architected a computer vision pipeline using MediaPipe and Django to turn recognition into a timed spelling game.
Diagnosed failure modes in commercial models and trained a high-fidelity custom classifier for the full alphabet.
Added temporal classification over sequential hand landmarks for dynamic gestures (e.g., J, Z).
Honours research on regional voting under spatially correlated noise: multi-scale ensembles, stability analysis, and a peer-reviewed IEEE UIC paper with oral presentation.
A real-time web version of the board game Stock Ticker: players authenticate with Google, game state lives in Firebase, and market data refreshes on a schedule via GitHub Actions/tasks.
Stock Ticker: live net worth, commodity chart, and market feed
Built the full client experience and realtime state flow as sole developer.
Used Google Authentication for player identity and session continuity.
Persisted data in Firebase with hourly updates driven by scheduled GitHub tasks.
A Flutter mobile/web helper that simplifies meal planning and grocery workflows. I led development, wired TheMealDB REST integrations, and kept user state persistent with Firebase and Google Authentication.
Led design and implementation of the Flutter + Firebase application.
Integrated TheMealDB APIs and persistent authenticated user state.
Shipped a playable web demo for quick sharing and demos.