Jul 2022 – Aug 2023 · 1 yr 1 mo · Bengaluru, Karnataka, India · Remote

• Developed viewpoint-agnostic Neural Networks by localising actions in both space & time to fit manufacturing use-cases
• Introduced vision transformers to correlate image features globally improving NNs’ longevity for on-premise deployments
• Multiplexed live video streams into single GPUs with 0% latency drops leading to 4x reduction in customers’ maintenance
• Owned end-to-end ML deployment for clients & delivered product POCs contributing to manufacturing efficiency improvement
• Designed deployment workflows of the ML arsenal from model intuitions and by analysing customer-specific requirements
• Migrated independent APIs into a consolidated MLOps Platform promoting Kubernetised training and deployment workflows

Jul 2021 – Jun 2022 · 11 mos · Bengaluru, Karnataka, India · Remote

• Ideated & Productized 4x light Action Recognition models maintaining same accuracy, reducing training time & costs by 90%
• Scaled annotation effort to 10% and increased NN’slifetime & generalisability by introducing Semi-Supervised technologies
• Designed heuristics & validation engines with beam search that postprocess raw Neural Network outputs into customer’s needs

Apr 2021 – Jul 2021 · 3 mos · Baleswar, Odisha, India

• Explored existing Object Detection and Tracking methodologies for the purpose of tracking Unmanned Aerial Vehicles
• Designed a Transformer-based architecture that directly predicts bounding boxes from images, developed in PyTorch
• Stabilized the confidence of prediction, which reduced latency of detection by 0.6 times and increased speed to 2.5 times
• Analysed and compared various object tracking algorithms including Kalman, Particle and Bayesian Filters with the detector
• Developed a motion-model over the filters to predict the movement of the object of interest in its future time-stamps

Jul 2020 – Jan 2021 · 6 mos · Kolkata, West Bengal, India

• Qualitative Research on Convolutional Neural Networks for Sequential Data associated with Natural Language Processing
• Experimented and analysed pre-existing research methodologies and publications related to Sign Language Recognition
• Explored Generative architectures and sequence-sequence models for the conversion of text input to Sign Language
• Developed a Sign Language Production architecture using Transformer-network to convert German Text input to German Sign Language using Python, in collaboration with the Coordinator of Centre for Artificial Intelligence and Machine Learning
• Analysed the computationally expensive strategy to develop a cheaper strategy in a smaller scale to reduce the generalized highly complex problem of Sign Language Production into a specific and computationally inexpensive problem