Artificial Intelligence

  • ​​I led Conversational AI R&D at Swiggy, India’s largest food ordering and delivery platform where I developed novel Deep Learning technologies for NLP and Speech. In NLP, I worked on Natural Language Understanding for chatbot, Intent as well as Sentiment Classification models. In Voice AI, I led research on Speech recognition for Hinglish, Voice sentiment analysis and led the team to publish Swiggy’s first AI paper.
 
  • I worked as a Research Scientist at Amazon Alexa AI in Seattle, USA, where I worked on advancing Alexa's Speech Recognition and Natural Language Understanding AI capabilities. I developed state-of-the-art end-to-end deep learning models for speech recognition using sequence models and distributed deep learning, and deployed high impact NLP models to detect and address sensitive and offensive conversations with Alexa for multiple languages using neural machine translation.
Deep learning for NLU & NLP
  • ​NLU modeling using weak supervision to decode intent in code-mixed chat
  • Sentiment analysis to identify negative feelings, emotions and opinions in chat 
  • Language identification in code-mixed chat 
  • Predicting Social Media Escalations based on chat input
  • Classification of Products into multiple categories based on text inputs
  • Named Entity Recognition of various entities in chat conversations
  • NLP for detecting sensitive and offensive chat interactions​
    • Trained deep learning NLP models to detect sensitive content in human-assistant interactions
    • Deployed NLP models for non-English languages using Neural Machine Translation and Text Generation techniques 
Deep learning for Speech
  • Developed state-of-the-art end-to-end deep learning ASR models based on sequence-to-sequence models
  • Significantly improved ASR model training times using Horovod for distributed training 
  • Trained ASR models on >20000 hours of data on AWS GPU instances
  • ASR modeling for "Hinglish", i.e. Hindi interspersed with English common amongst Indian speakers
  • Spoken Language Identification for low-resource Indian languages code-mixed with English 
  • Speaker Diarization modeling, i.e. "who spoke when"  for code-mixed speech 
  • Audio Sentiment Analysis to understand emotions expressed in speech
Privacy-preserving deep learning
  • Homomorphic encryption, federated learning and its applications for private machine and deep learning​
Papers

Gupta A, Sukumaran R, John K, Teki S (2021)
Hostility Detection and Covid-19 Fake News Detection in Social Media
CONSTRAINT Workshop, AAAI 2021 (non-archival) [download]
​
Vijjali R, Potluri P, Iyer S, Teki S (2020)
Two Stage Transformer model for Covid-19 Fake News Detection and Fact Checking 
NLP for Internet Freedom Workshop co-located at COLING 2020 [download] [data]

Rangan P*, Teki S*, et al. (2020) [*Equal contribution as first authors]
Exploiting Spectral Augmentation for Code-Switched Spoken Language Identification 
1st Workshop on Spoken Language Technologies for Multilingual Communities, INTERSPEECH 2020 [download]

Brahma AK, Potluri P, Kanapaneni M, Prabhu S, Teki S (2021)
Identification of Food Quality Descriptors in Customer Chat Conversations using Named Entity Recognition 
Accepted at CODS-COMAD 2021 Research Track. [download]
Blogs
​