Feb 2024 – Jun 2024 · 4 mos · Remote

• 1. Innovated an evaluation method to measure geographical biases in text-to-image models like Stable Diffusion and DALLE, enhancing precision and diversity calculations using decoupled object and background representations in images
• 2. Implemented Segment-Anything model and zero-shot text to bounding box techniques to decouple image elements
• 3. Developed PatchViT, a novel technique that selects relevant image patches for enhanced feature extraction in Vision Transformer (ViT) models, outperforming standard ViT and CNN-based methods
• 4. Analysed metric trends across regions, enhancing understanding of model performance under various prompt settings
• 5. Applied CLIP Zero Shot classification to real \& generated datasets, assessing potential biases from regional information

Apr 2022 – Jun 2023 · 1 yr 2 mos · Mumbai, Maharashtra, India · Hybrid

• 1. Developed a Vision Transformer model using PyTorch and HuggingFace for multilingual handwritten document recognition, resulting in a 40% improvement in recognition rates.
• 2. Engineered a synthetic dataset generation algorithm to enhance domain adaptation and model generalization, leading to a 50% performance boost across six benchmark OCR datasets.
• 3. Integrated and fine-tuned LayoutLM and TableNet models to optimize document understanding and table detection, achieving a 30% increase in layout and text recognition efficiency.
• 4. Streamlined a semi-supervised data programming pipeline, allowing programmatically labeled training data using user-defined labeling functions and rules, eliminating the need for manual labeling.
• 5. Deployed end-to-end APIs for benchmarking and frontend integration using FastAPI, Postman, and Flask.
• 6. Authored a paper introducing a novel Priority-Based Suppression technique on a combination of layout models for digitization of multilingual scanned documents.
• 7. Instructed a Udemy course titled "A Hands-On Introduction to OCR" with over 800 enrollments.

Jan 2022 – Jun 2022 · 5 mos · Mumbai, Maharashtra, India

• 1. Created an Enhanced EfficientNet (EEF-Net) model for Breast tumor detection with 97% accuracy and 98% F1-Score on mammography scans.
• 2. Devised a pre-processing pipeline to remove the pectoral muscle from the mammography scans using Contrast Limited Adaptive Histogram Equalization and K-Means Clustering, boosting the model accuracy by 17%.
• 3. Conducted a comprehensive analysis and comparison of the current state-of-the-art models with EEF-Net, resulting in the publication of a paper in Elsevier SSRN.

Jun 2021 – Dec 2021 · 6 mos · Pune, Maharashtra, India · On-site

• 1. Headed a state government project focused on automating license plate detection of non-helmeted motorcycle riders, which resulted in a 25% reduction in traffic law violations at over 10 signal junctions in Pune.
• 2. Implemented a three-stage system incorporating the YOLO model for motorcycle, helmet, and license plate detection, achieving a Mean Average Precision score of 99%.
• 3. Optimized and integrated the model into the hardware system using TensorFlow Lite and Google Cloud, yielding a 70% reduction in real-time response time.