Jatin Kulkarni

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

• Reduced median top-1 prediction error by over 1,600 km against the CLIP baseline
• Increased predictions within 750 km by nearly four times
• Evaluated embedding strategies and geographic-resolution tradeoffs

Compared multimodal and image-only approaches for diabetic retinopathy classification, including analysis of semantic alignment and fine-grained visual discrimination.

• ResNet-50 achieved 92.52% accuracy
• The two-stage CLIP approach achieved 89.31% accuracy
• Used Grad-CAM and t-SNE to examine model behavior and embedding structure

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

• ViT-Base reached 72.87% symbol-recognition accuracy
• The sequence model achieved 66.19% exact match
• Achieved a BLEU score of 0.8428

Built a full-stack application for answering natural-language questions over financial PDFs using retrieval-augmented generation.

• Implemented context-grounded retrieval with SentenceTransformers and FAISS
• Built the backend API with Python and FastAPI
• Developed the user workflow with Angular and Tailwind CSS

Combined CNN-based spectrogram analysis with RNN and LSTM sequence modeling to classify music genres from audio.

Built an interactive web experience that turns design-thinking methods into guided problem-solving steps.