Designed a diffusion-hypernetwork framework with noise and weight consistency losses to reduce (not just quantify) aleatoric/epistemic uncertainty; delivered sharper, more stable CT/MRI reconstructions without sacrificing PSNR/SSIM.
Demonstrated measurable gains in clinical tasks: +0.16 Dice on CT lung-nodule segmentation under distribution shift and AUC/accuracy improvements on fastMRI Knee and Brain, especially when using uncertainty maps as reliability inputs.
Built AUFairGAN, a GAN-based super-resolution model with uncertainty-aware discriminator and tailored loss to mitigate aleatoric-uncertainty bias across gender and race in facial age estimation under distribution shift.
Mitigated aleatoric-uncertainty disparities on CelebA, FairFace, and UTKFace, improving accuracy by up to +3.5% across groups, while maintaining image quality (PSNR/SSIM).
Built a transformer-based cross-domain recommender system transferring knowledge from books to movies to generate serendipitous recommendations.
Integrated GPT 3.5 Turbo and Flan-T5 for cross-domain representation learning, improving recommendation accuracy and relevance.
Pulled COVID-related tweets/comments from 3B AWS records via SQL; optimized costs to <$100 and quantified topic-alignment trends between CDC communications and public responses.
Quantified uncertainty in CNN/BERT on MovieLens and Amazon Books and implemented data minimization that cut required data by 25% while maintaining recommendation accuracy and strengthening user privacy.
Github Publication Presentation
Processed 20k unstructured threat reports; built an NLP pipeline to extract malicious actions and auto-map to MITRE ATT&CK, achieving 74% mapping accuracy and reducing manual triage for Bank of America analysts.
Built a pipeline to ingest and preprocess Twitter data and learn trend-surfacing policies with RL (bandit-style reward from engagement/topic momentum).
Identified emerging product topics earlier than simple keyword tracking and generated ranked promote recommendations for downstream decision-making.
Developed a Flask web app for organizing car-enthusiast meetups with user accounts, event creation/RSVP, and search/browse by time and location.
Implemented a RESTful backend with templates and forms, basic CRUD workflows, and a relational database for persistent event/user data.
Built a full-stack web app in Django for listing management, user auth, and admin workflows.
Implemented a client–server chat application using Java sockets.
Wrote a basic lexer/parser to compile simple arithmetic programs (e.g., add, sub).
Developed a scoring tool to record runs/wickets and compute basic match stats.
Programmed puzzle/board-game solvers in Java using search and constraint techniques.
Implemented Convex Hull, Voronoi Diagram, and Polygon Triangulation algorithms.