Mar 2020 – Jan 2021 · 10 mos

• Worked on Samsung’s Voice Assistant, Bixby. Early adopted, wrote, and trained transformer models (BERT, etc.) on millions of user queries for domain and intent prediction tasks. Handled the ML pipeline - regular training, evaluation, and iterative model release.
• Utilized ML methods such as label smoothing, incremental learning to boost model performance, and knowledge distillation to deploy models on-device. Migrated domain and intent from a rule based engine to a smart semantic system.

Jul 2018 – Mar 2020 · 1 yr 8 mos

• Worked to implement a open domain QnA chatbots inside Samsung's voice assistant using a multi-stream deep learning architecture. Leveraged Graph Convolution Networks, Memory Networks, knowledge graphs, and BERT. Work published at LREC 2020.

May 2018 – Jul 2018 · 2 mos · Tübingen, Baden-Württemberg, Germany

• Worked at the intersection of Neuroscience and Deep Learning. Project aimed at estimating the physical properties of clothes or fabric (stiffness, elasticity, firmness, etc.) from its varying videos. Experimented with CNN+LSTM networks, 3D CNNs, 2 stream, multi loss networks along with the optical flow for video feature extraction. Proposed triplet loss based optimization for model training

May 2017 – Jul 2017 · 2 mos · Bengaluru, Karnataka, India

• Goal was to identify confidential text inside document images (credit cards etc.) and build an image classifier to judge the degree of risk. Experimented with early versions of OCR technologies. Also, used RNNs to model sequential text image features.
• Trained a deep learning model, built an android application to deploy the model on edge device.