Work Experience


  • Amazon | Applied Scientist Jul'23 – Present
    Last Mile Team | Manager: Saurabh Sohoney | Hyderabad, India
    • Proposed an iterative k-core graph decomposition algorithm to improve clustering in an address catalog creation pipeline; work published at ACM SIGSPATIAL 2024.
    • Refactored and modularized the pipeline (spanning named entity recognition, entity matching, graph clustering) into a scalable, object-oriented codebase, consolidating previously isolated geo-specific workflows.
    • Productionized the system with ~3,000 lines of PySpark/PyTorch code deployed across multiple geographies; large-scale A/B testing demonstrated 1.5% gains in package consolidation efficiency translating to increased deliveries per hour.

  • Amazon | Applied Scientist Intern Jan'22 – Jun'22
    Last Mile Team | Manager: Saurabh Sohoney | Hyderabad, India
    • Adapted and customized weak supervision techniques (e.g. Snorkel, Dawid-Skene) to reduce manual annotation effort for core internal problems in the team.
    • Achieved within 5% of fully supervised model performance on two tasks — address text matching & GPS activity classification.
    • Among the team's first explorations in weak supervision; submitted the work to Amazon's internal ML conference and my implementation was later adopted by other members.

  • Indian Institute of Technology, Hyderabad | Research Assistant Apr'20 – Jul'23
    Machine Learning & Vision Group (Lab 1055) | Guide: Prof. Vineeth N Balasubramanian
    • Developed algorithms for learning with limited supervision under domain shift, published two papers.
    • Introduced the novel problem setting of Unsupervised Domain Generalization (UDG) and showed contrastive self-supervision as a strong baseline – NeurIPS 2021 workshop, PMLR v181 (2022).
    • Proposed a simple test-time adaptation method for source-free domain generalization – ICLR 2023 workshop.

  • Preferred Networks | Research Intern Jul'19 – Sep'19
    Sports Team | Guides: Alexis Vallet, Daichi Suzuo | Tokyo, Japan
    • Proposed a player performance evaluation method in football by modeling the spatial context of all players using tracking data.
    • Designed trajectory- and position-based features to improve estimation of per-action goal-scoring value; achieved 35% precision gain over baselines.
    • Work resulted in a Japanese patent filing and a technical blog on the company website.

  • International Institute of Information Technology, Hyderabad | Research Assistant Jul'18 – Jun'19
    CVIT | Guide: Prof. C. V. Jawahar
    • Explored ball detection in broadcast tennis footage, tackling camera motion, occlusion, and visibility issues.
    • Created and annotated a 1000+ frame dataset across varied match conditions; experimented with multiple models, finalizing on Faster R-CNN, and explored preliminary ball trajectory prediction methods.

  • Samsung Research | Summer Intern May'18 – Jul'18
    Vision Intelligence Group | Guides: Rabbani Patan, Gaurav Kumar Jain | Bangalore, India
    • Developed a scene-based content recommendation approach inspired by Word2Vec; evaluated on the ADE20K dataset.
    • Internship offered as a result of winning the Samsung Bixby Hackathon held at IIT (ISM) Dhanbad.

  • Indian Statistical Institute, Kolkata | Research Intern Nov'17 – Dec'17
    Machine Intelligence Unit | Guide: Prof. Ashish Ghosh
    • Proposed a novel outlier detection framework using the projection principles of stacked auto-encoders and probabilistic neural networks.
    • Paper accepted in the journal of Pattern Recognition.

  • Samsung Research | Summer Intern May'17 – Jul'17
    Web Team | Guide: Amit Sarkar | Bangalore, India
    • Worked on an Abstractive Text Summarization scheme for generating link previews of websites for the Samsung Mobile Browser.
    • Formulated a model based on Attention Based Encoder-Decoder RNN for summarization and trained and tested it on the Insight BBC Dataset.