Hi I’m Ishan Kumar, a second year MS CS student at Columbia University, pursuing the Machine Learning Track. Previously I was a Data and Applied Scientist at Microsoft India. I worked in the Search Technology Center, India as part of the Segments Team of Bing for 2 Years. I am looking for Full time roles starting Jan '26, if your team is hiring for MLE, Applied Research roles please reach out!

During the summer of 2025, I am interning at the JP Morgan Machine Learning Center of Excellence as an Applied Science Intern, where I will be diving deep into LLM Evals and building better systems to evaluate the internal RAG and QA pipelines.

I also work as a part-time research assistant in Professor Goldblum’s lab, where we are exploring the unique structure of Diffusion Language Models to speed up their inference.

Prior to this, I was a part-time researcher in the Natural Language Processing Lab at ETH, working on Causality and NLP under the supervision of Zhijing Jin, Prof. Mrinmaya Sachan, and Prof. Bernhard Schölkopf. Our work CausalCite was published in ACL 2024.

I obtained my Bachelor’s Degree at the Indian Institute of Technology, Roorkee during which I was also a Research Student in the Molecular Simulations and Drug Delivery Research Group in IIT Roorkee with Prof. Prateek Jha. Previously, I have also interned with the Video Analytics Lab, IISc with Prof. Venkatesh Babu.

I read that over 90% of ML models never end up getting deployed, and I decided I want to change that. I am passionate about improving inference time optimisations for LLMs to improve their deployability (dLLMs, which have the potential to output >1 tokens at a time, is one way of doing that—but I love to explore different directions like MQAdapt). I am also interested in leveraging Gen AI to solve real-world user problems and drive solutions end to end. Please reach out if you are interested in working together!

💻 Experiences

  • Research Assistant, ETH Zurich and MPI
    Worked on CausalCite: a causal formulation of paper citations. Accepted in ACL 2024, presented as a first author in Bangkok.
  • Data and Applied Scientist, Microsoft
    Worked on the segments team, deployed pipelines which resulted in a DAU increase from 500k to 1.5M for tech queries. Published SegRank: a novel grounding method for LLMs.
  • Research Intern, Deloitte AI Center of Excellence
    Created an internal MLOps Platform
  • ML Engineering Intern, Zomato
    Fraud Detection for new users using ML
  • Research Student, Video Analytics Lab IISc Banglore
    Created a self supervised pipeline to get Pose and Colored 3D mesh from a 2D RGB image.
  • Research Intern, Complex Networks Research Group (CNeRG) IIT Kharagpur
    Made a Question Answering model for AmazonQA dataset

    📝 Publications

  • Ishan Kumar, Zhijing Jin, Ehsan Mokhtarian, Siyuan Guo, Yuen Chen, Negar Kiyavash, Mrinmaya Sachan, Bernhard Schoelkopf : CausalCite: A Causal Formulation of Paper Citations; Accepted at ACL Findings 2024. ArXiv GitHub
  • Ishan Kumar, Diptiman Purbey : SegRank: A Novel Approach to Creating Grounding Data For LLMs, Microsoft Journal of Applied Research 2023
  • Ishan Kumar, Prateek Jha : Coarse‑Grained Configurational Polymer Fingerprints for Property Prediction using Machine Learning; ArXiv GitHub
  • Ishan Kumar, Shikhar Saxena : Improving Mini‑Bert’s Predictions on COVID Fake News using Multilingual Knowledge Distillation; AI4SG workshop at IJCAI 2021.

    I like to contribute to Open Source, in the past I have actively contributed to PyTorch Ignite PyTorch-Ignite (one of their Top-10 Contributors!), which is a High-level library to help with training and evaluating neural networks.

    During the period of April to Nov of 2020, I worked with Video Analytics Lab, IISc on a task of Self Supervised Hand Mesh Detection and Color Recovery. The purpose of doing this in a self supervised manner was to use the large number of unlabelled data available on the internet and makes the model more robust since its trained on real world outdoors data also.

    In the summer of 2020, I worked with CNeRG Group, IIT Kharagpur on a task of Question Answering on the AmazonQA Dataset. It was a Reading Comprehension task where the the questions had to be answered using the reviews as context.

    Interests

    Natural Language Processing, Inference time optimisations for LLMs which don't require further training. Deploying Machine Learning Models for user needs.

    If you have any questions or would like to collaborate, feel free to reach out to me through email ishankumar216@gmail.com