Buffalo, New York

• Created a text+pose guided image generation system using latent diffusion models (ControlNet with Stable diffusion weights).
• Added noise and performed latent interpolation on best pair human face images to generate high-resolution morphs.
• Generated images at different noise levels and selected the best candidate by computing CLIP similarity scores for fixed prompts.
• Developed an image manipulation detection system to detect, localize, and label tampered parts in news articles.
• Trained a CNN model with a Resnet-50 backbone to detect jpeg compression errors in tampered images with 93% accuracy.
• Utilized NEDB-Net to extract noise and edge-based features to localize and get manipulation masks for the tampered regions.
• Fine-tuned a custom Yolo-v8 model to detect objects in the localized regions and classify them into 18 categories.
• Created analytic containers to detect image text inconsistencies in multimodal news articles.
• Developed a text-transformer tool to perform controlled named entity and parts of speech replacements in texts using SpaCy.
• Extracted text from images using Paddle OCR and developed contradiction and mismatch detection algorithms to detect inconsistencies between image text and news body.

Bhubaneswar, India

• Trained random forest model to classify dependent and independent parts for refrigeration systems with 97% accuracy.
• Developed a module to predict part prices using XGBoost regressor and integrated with CUMMINS part pricing system.
• Built API service using Django REST Framework to parse delegation data from files received in hourly batches.
• Developed employee task delegation system to manage production workflows and approvals using python.
• Reduced batch data transfer failure rate by 20% through automation and monitoring of batch jobs using RPA and Appworx.
• Managed deployment and maintenance pipelines for system availability and reliability using Jenkins and Git/Github.