Projects.

Replaced the flat VQ codebook of the one-step diffusion restorer OSDFace with a hierarchical region-aware quantizer that parses faces with BiSeNet and quantizes each region against its own three-level residual codebook.

Cut FID by roughly 30% on synthetic CelebA and real-world LFW benchmarks at under 5% added inference cost, and validated the learned codes through region-level code-swap interventions.

Extended the Lookback Lens hallucination detector with per-token attention, entropy, and logit features from LLaMA-2-7B-Chat, replacing span-level averaging with 1D-CNN and BiLSTM temporal probes.

Raised detection AUROC from 0.833 to 0.885 on AggreFact, with cross-domain transfer measured on XSum and NQ-Open.

Engineered a dual-branch vision-language model with frequency-aware spatial refinement for prompt-driven medical segmentation.

Improved IoU from 0.490 to 0.882 and Dice from 0.505 to 0.923 over the CLIPSeg baseline on ETIS polyps.

Analyzed MBTA bus performance to identify service equity gaps by correlating historical ridership and reliability with community demographics.

Prototyped time-series and tree-based models to forecast delays and isolate drivers of poor reliability for underserved neighborhoods.

Digitized ECG paper records into time-series signals using an OpenCV pipeline with thresholding and contrast enhancement.

Trained a hybrid CNN-LSTM on PTB-XL, reaching roughly 98.87% accuracy on cardiac abnormality classification.

Implemented DCGAN, Conditional GAN, and WGAN in TensorFlow and Keras for data augmentation on limited datasets.

Used SSIM and PSNR to compare fidelity and selected DCGAN for its stronger feature realism in this setting.