Machine Learning Engineer II, NYU Langone Health · New York, NY
I build and evaluate machine learning systems. Most of my recent work has been in healthcare and applied research: medical imaging, clinical prediction, and the evaluation of language models.
I'm currently a Machine Learning Engineer II at NYU Langone Health, where I assess the technical and clinical feasibility of AI/ML approaches across the health system and build the models and data pipelines used to validate them before anything scales. I hold an MS in Computer Science from NYU's Courant Institute, with an AI concentration, and a BTech in Computer Science from IIT Mandi.
Recent projects have been in medical imaging, LLM evaluation, and self-supervised representation learning.
GNews-Gemini, a dataset of LLM-generated disinformation spanning five manipulation strategies, tested against ISOT by cross-dataset transfer. Detectors scoring 99.3% and 96.9% within their own data fell to 48.3% and 51.5% across the authorship boundary. An authorship classifier reached 99.7%, indicating the models were measuring who wrote the text rather than whether it was true.
Can a non-invasive MRI scan stand in for a biopsy? Kidney cancer patients were imaged before partial nephrectomy, and IVIM parameters plus their radiomics were correlated against both pathologist scoring and quantitative multiplex immunofluorescence from the resected tissue. Tissue diffusivity tracked cellular density; pseudo-diffusivity tracked vascularity.
Most scientific figures are vector graphics, so the usual image-extraction approach silently misses them. This pipeline uses a layout model instead, links each figure and table to the body paragraphs that discuss it, and returns structured JSON per item. 320/320 tables and 496/540 figures across 100 papers from 10 publishers.
A recurrent Joint-Embedding Predictive Architecture trained on 2.5M frames of two-room navigation, learning structured state representations with no pixel reconstruction at all. VICReg regularisation kept it from collapsing to a constant; final agent position error was 1.89 MSE.
A Flask and PostgreSQL REST service with full CRUD, condition filtering and restock endpoints, built test-first to 95%+ coverage across unit and BDD suites. Shipped through a Tekton pipeline onto OpenShift, and deployed to a local K3s cluster as a three-replica service behind a load-balanced route. Team project for NYU's DevOps and Agile Methodologies course.
Fake News Detection Across the Human-AI Divide
International Conference on Big Data Analytics (ICBDA), 2026
Cross-validation of intravoxel incoherent motion metrics with quantitative and qualitative histology in excised kidney tumors
International Society for Magnetic Resonance in Medicine (ISMRM), 2026
A diffusion weighted MR study of individual kidney function decline in kidney cancer patients
International Society for Magnetic Resonance in Medicine (ISMRM), 2026
Pre-surgical DTI and IVIM metrics, acquired across bipolar and flow-compensated sequences at systolic and diastolic gating, were tested against the one-year change in individual kidney measured GFR. Twenty-three parameters correlated significantly with functional decline.
NYU Langone Health · New York, NY
Predictive Analytics and AI Research Lab, NYU Courant · Advisor: Prof. Anasse Bari
NYU Langone Health, CAI2R · PI: Dr. Eric Sigmund
Tata Innovation Labs, Delhi · Mentor: Pankaj Malhotra
New York University · Courant Institute of Mathematical Sciences
Indian Institute of Technology (IIT) Mandi
Python · SQL · C++ · R · JavaScript · MATLAB
PyTorch · TensorFlow · Keras · scikit-learn · XGBoost · BERT · NLP
Hugging Face Transformers · RAG · FAISS · fine-tuning · LLM APIs · LangChain
NumPy · Pandas · SciPy · NLTK · spaCy · Spark · OpenCV · Matplotlib · seaborn · SPSS · hypothesis testing
PostgreSQL · Docker · Kubernetes · AWS / GCP · REST APIs · Git · Tekton CI/CD · HPC
Tableau · Jupyter · Flask · OpenShift · pytest · Behave / Selenium
Email is the surest way to reach me.