CV

The PDF icon on the right opens the two-page version.

General Information

Full Name Kawshik Manikantan Sundar
Position Pre-Doctoral Researcher, Languages Group, Google DeepMind
Location Bengaluru, India
Email skawshik11@gmail.com
Profiles Google Scholar · GitHub · LinkedIn

Education

  • 2019 – 2025
    B.Tech. + M.S. by Research (dual degree), Computer Science
    International Institute of Information Technology, Hyderabad (IIITH), India
    • CGPA: 9.27/10; Dean's List (top 10–15%) in five consecutive semesters (Spring 2021 – Spring 2023).
    • Thesis: Coreference without Bells and Whistles; advisors: Vineet Gandhi and Makarand Tapaswi.
    • Coursework
      • AI: Advanced NLP, Computer Vision, Statistical Methods in AI, Advanced Graph Learning, Information Retrieval and Extraction, Digital Image Processing.
      • Systems: Data Structures and Algorithms, Operating Systems and Networks, Design and Analysis of Software Systems.

Research Experience

  • 2025 – now

    Bengaluru, India

    Pre-Doctoral Researcher
    Google DeepMind, Languages Group
    • Developing agentic training methods for efficient deep reasoning over long videos.
    • Post-training Gemini to improve video reasoning across multilingual and multicultural contexts; co-authored the MINERVA-Cultural benchmark (CVPR 2026, Highlight).
    • White-box analysis of cross-lingual factual retrieval in Gemini.
  • 2024 – 2025

    Bengaluru, India

    Research Fellow Intern
    Advanced Micro Devices (AMD), AI Models Group
    • Improved generalization accuracy of a continual model-editing pipeline by ~15% with multi-faceted data augmentation.
    • Built a continually updated, large-scale benchmark of recent Wikipedia edits to measure how well edited models adapt to new facts.
  • 2021 – 2025

    Hyderabad, India

    Undergraduate Researcher
    Center for Visual Information Technology (CVIT), IIITH
    • Proposed Major Entity Identification (MEI), a generalizable alternative to coreference resolution that transfers across domains and enables LLM evaluation (EMNLP 2024).
    • Built IdentifyMe, a long-context mention-resolution benchmark for evaluating frontier LLMs (NAACL 2025, Oral).
    • Multi-view detection and tracking in long sports videos, in collaboration with Star Sports.
  • 2022

    Remote, India

    Research Intern
    Star Sports
    • Built and deployed an operator-free player-tracking system used live during the Asia Cup 2022 broadcasts.
    • Drove ideation and feature development, and built the pre-deployment pipeline.

Publications

  • 2026
    MINERVA-Cultural: A Benchmark for Cultural and Multilingual Long Video Reasoning
    D. Singh, A. Nagrani, K. Manikantan, H. Singh, D. Tewari, T. Weyand, C. Schmid, A. Angelova, S. Dave. CVPR 2026 (Highlight).
    arXiv
    data
  • 2025
    IdentifyMe: A Challenging Long-Context Mention Resolution Benchmark for LLMs
    K. Manikantan, M. Tapaswi, V. Gandhi, S. Toshniwal. NAACL 2025 (Oral).
    paper
    arXiv
    code
  • 2024
    Major Entity Identification: A Generalizable Alternative to Coreference Resolution
    K. Manikantan, S. Toshniwal, M. Tapaswi, V. Gandhi. EMNLP 2024.
    paper
    arXiv
    code
    project page

Honors and Awards

  • 2026
    • CVPR 2026 Highlight for MINERVA-Cultural.
  • 2025
    • Oral presentation at NAACL 2025 for IdentifyMe.
  • 2021 – 2023
    • Dean's Merit List, IIIT Hyderabad, five consecutive semesters (Spring 2021 – Spring 2023).

Teaching and Service

  • Reviewer: 7th Workshop on Narrative Understanding (WNU), NAACL 2025.
  • Teaching Assistant, Statistical Methods in AI, IIITH (Fall 2023): designed assignments, graded coursework, and conducted tutorials.
  • Mentoring and organizing: mentored incoming first-year students at IIITH; organizing team of Felicity, the institute's techno-cultural fest.

Selected Projects

  • A Study of Tracking
    Re-implemented Tracking at 100 FPS (GOTURN) and the neural MOT solver (MPNTrack), and extended the latter towards multi-view tracking. PyTorch, PyTorch Geometric.
    github
  • Modeling for NLP Tasks
    Compared Bag-of-Words, LSTM, and DIIN models for natural language inference; implemented character-level convolutional (CREPE) and attention-based (CANINE) text classifiers; multi-class classification of mental-health disorders from Reddit posts with CNN, Random Forest, SVM, XGBoost, and DistilBERT. PyTorch Lightning, Keras, NLTK.
    NLI
    CREPE
    mental health
  • SimpleRA DBMS
    Relational DBMS with integer tables and matrix operations: two-phase on-disk merge sort, delayed-write buffer manager, linear-hash and B+-tree indexes. C++.
    github
  • Wikipedia Search Engine
    Search engine over a full Wikipedia dump, built from scratch: field-aware inverted index (title, infobox, body, categories, links, references) built by external k-way merging of partial indexes, with stemming, stop-word filtering, and field-restricted queries. Python, NLTK.
    github
  • Linux Mini Shell
    Shell with piping, redirection, background processes, and common built-ins, implemented directly on POSIX system calls. C.
    github

Skills

  • Programming
    • Python, C/C++, Java, JavaScript, MATLAB
  • ML and NLP
    • PyTorch (Lightning, Geometric), TensorFlow/Keras, scikit-learn, XGBoost, Hugging Face Transformers and Datasets, NLTK
  • Tools
    • Linux, Git/Mercurial, Bazel