Jugal R Sheth Email & Phone Number

Menlo Park, CA, US

Voice AI Assistant for wearables

New York, NY, US

• Led a team to develop an MLOps framework based on Kubeflow Pipelines for continuous training, deployment, serving, and model monitoring, resulting in faster time-to-market and increased model accuracy.
• Built a two-tier candidate generation and ranking system for information retrieval and entity matching using transformer based language models
• Built frameworks for representation learning using Transformers, CNNs, LSTMs architectures for few-shot learning on highly dynamic data
• Built scalable ETL pipelines using Airflow and Apache Beam extracting features from over half a billion data points
• Developed a high-performance nearest neighbor search system on over half a billion embeddings using tools such as FAISS and the Vertex Matching Engine (scaNN), resulting in faster inference times and improved search.
• Successfully served language models across large-scale distributed GPUs using TensorFlow serving, ensuring efficient and reliable inference.

Seattle, WA, US

• Developed object and human detection systems for ecommerce applications in PyTorch
• Developed video detection and object tracking systems using 2D and 3D convolutional neural networks. Used Hungarian algorithm to improve tracking performance
• Developed human-pose estimation system with Stacked hourglass in TensorFlow
• Trained a feature extraction model for visual similarity search using Triplet loss and Neighborhood component analysis (NCA) in PyTorch
• Trained instance and semantic segmentation models for visual search applications
• Created a system for gender, age, fine-grained clothing attributes tagging from videos and images using attention mechanisms

Developed an algorithm for weighted nearest neighbor classifier using a variation of softmax classifier as a discriminant function with second order Newton’s algorithm for optimal learning factor for faster convergenceTrained text classification and sentiment analysis models using bidirectional recurrent neural networks and GloVe Embeddings in.

Eskilstuna, SE

Developed applications for real time pedestrian detection using stereo camerasApplied Semi-Global Matching to generate disparity maps and dense depth map for precise distance of pedestrians from the cameraImplemented Aggregate Channel Features for feature extraction and Adaboost classifier for pedestrian classificationAdded Kalman Filter to the system to.

Lohr Am Main, DE

Hands-on training on Hydraulics, Pneumatic systems.Hydraulic, pneumatic sequencing circuit building. Bosch Rexroth PLC Programming using Indralogic software. Worked on different types of sensors

Master’S Degree, Electrical And Electronics Engineering

B-Tech, Electrical And Electronics Engineering