Dynamic Object Classification• Developed, tested, and deployed sequential model for dynamic object tracking, enhancing system accuracy by 8%.• Engineered a robust data pipeline capable of processing over 4000 hours of in-house data, significantly improving model generalization .• Handled long-tailed data distributions of data, achieving a 4% reduction in validation error.• Implemented Quantization Aware Training (QAT) to optimize the model, reducing inference time by… Show more Dynamic Object Classification• Developed, tested, and deployed sequential model for dynamic object tracking, enhancing system accuracy by 8%.• Engineered a robust data pipeline capable of processing over 4000 hours of in-house data, significantly improving model generalization .• Handled long-tailed data distributions of data, achieving a 4% reduction in validation error.• Implemented Quantization Aware Training (QAT) to optimize the model, reducing inference time by 20%,decreasing the model size by one-third, and generating an ONNX model to enhance real-time applicationperformance.• Leveraged ML workflows for efficient training and validation of deep learning models.• Conducted multiple critical code reviews, elevating overall code quality and maintaining best practices• Collaborated closely with cross-functional teams, including product owners, to translate complex business problems into actionable data science solutionsImproving Code-LLM Accuracy on Buggy Code Completion• Curated an internal dataset with 1,000+ buggy code examples to test Code-LLMs in realistic scenarios.• Identified a 50%+ drop in model pass rates in buggy contexts, revealing key robustness gaps.• Applied bug-mitigation strategies like removal-then-completion and rewriting-then-completion, improving model accuracy by 20%. Show less

• Developed and deployed a high-performance 3D object detection model optimized for on-device applications, improving detection accuracy and responsiveness.• Leveraged a diverse range of datasets for comprehensive training and validation, achieving a 6% Mean Average Precision (MAP) improvement over existing models.• Implemented model optimizations using ONNX and TensorRT, resulting in a 20% reduction in inference time, thereby enhancing real-time application performance

• Implemented SOTA models for text and vision domains, effectively handling low-data regimes and maintaining high accuracy.• Engineered a robust architecture that outperformed existing models by 9% in the text domain under low-data conditions.• Set new performance benchmarks and contributed to academic research, resulting in a publication at the IEEE International Joint Conference on Neural Networks (IJCNN) 2023.