Work

Machine Learning Engineer 2

Comcast, Chennai

Jan 2023 - Present

  • Worked on experimenting with multiple hardware and multiple backend REST APIs to find the efficient configuration to deploy a transformer encoder model for the Xfinity Assistant
  • Experimented with Kubernetes pods/cpus to find the impact on the model's latency and throughput
  • Created a POC of a LLM based chatbot with an external database(RAG) for answering technicians' queries

    • Data Science Consultant Intern - Part Time

    Tata Communications

    August 2022 - October 2022

  • Worked on creating new features on the application for Wireless and Wireline Network Expansion and Planning
  • Reduced the latency of the Streamlit app with various optimizations.

    • Research Intern

    National University of Ireland, Galway

    Nov 2021 - April 2022

  • Proposed Adapter based efficient Transformers for Offensive Language Detection for low resource and codemixed languages.
  • Developed a Multimodal misogyny meme identification system using late fusion with CLIP and transformer model
  • Volunteering a shared task on ”Emotional Recognition in Tamil” for the DravidianLangTech workshop at ACL2022

    • Applied Research Intern - Low Resource NLP

    NVIDIA - Bangalore

    Dec 2021 - March 2022

  • Created Monolingual corpora for four under resourced languages of about 25 GB each from existing open source corpora
  • Developed 345 M GPT 2 models for South Indian low resource languages using Megatron-LM and analyzed their perfomance in downstream tasks with the other Multilingual models

    • Software Engineering Intern(AI/ML Team)

    Impiger Tech

    May 2021 - Nov 2021

  • Primarily worked on Invoice extraction system, learned about common OCR tools such as tesseract, Camelot, and ocrmypdf
  • Researched existing techniques on invoice automation and employed an object detection-based approach which is both efficient and involves less cost in annotation(other methods require commercial OCR tools as annotation). Learned and employed the YOLO v5 state-of-the-art model in this process.
  • Analyzed transformer-based models for handwritten text extraction
  • Created rule-based methods for signature and seal detection
  • Implemented state-of-the-art pretrained models as activities in ImpigerRPA framework, learned on productionizing ML models as web services via Flask.
  • Formulated an approach for volunteer and senior citizen matching using NLP techniques