Built a ViCLIP-based geolocation system that learns richer geospatial embeddings from multi-frame panoramic imagery and captions.

• Reduced median top-1 prediction error by over 1,600 km vs CLIP
• Increased predictions within 750 km by nearly 4x
• Improved top-1 retrieval quality across distance thresholds

Explored multimodal and image-only approaches for diabetic retinopathy classification, comparing semantic alignment against fine-grained visual discrimination.

• ResNet-50 achieved 92.52% accuracy
• Two-stage CLIP approach achieved 89.31% accuracy
• Used Grad-CAM and t-SNE to analyze explainability and embedding structure

Software engineering work on SageMaker Training Plans infrastructure for reserved-capacity procurement, allocation, and validation across ML training and inference workloads.

• Reduced deployment/configuration turnaround from 6–10 days to 1–2 days
• Reduced customer friction by ~96% through data-driven reserved-capacity limit redesign
• Contributed to inference-related Training Plans workflows for reserved GPU/accelerator capacity

Built a multi-stage pipeline to convert handwritten equations into LaTeX using visual recognition and sequence generation models.

• ViT-Base reached 72.87% symbol recognition accuracy
• Seq2Seq model achieved 66.19% exact match
• BLEU score of 0.8428

Built a full-stack application for answering natural language questions over financial PDFs using FastAPI, Angular, FAISS, and Hugging Face models.

Predicted music genres by combining CNN-based spectrogram understanding with RNN/LSTM sequence modeling for richer audio representation.

Built an interactive web experience that turns design thinking workflows into guided, game-like problem-solving steps.