At Sanofi I work deep in the science — building survival and longitudinal models that track how diseases progress through time, drawing from rich multimodal datasets: cognitive endpoints, MRI/PET neuroimaging, genomic biomarkers. Every model has to hold up under scrutiny — I evaluate discrimination, calibration, and subgroup robustness before anything reaches a trial team. The pipelines I build follow CDISC standards (SDTM/ADaM) end-to-end, with automated validation and full regulatory traceability. Working shoulder-to-shoulder with Biostatistics and Clinical Development, I translate messy clinical questions into model specs that are actually useful.

Four years at Cognizant meant wearing many hats across pharma and life sciences. I built patient stratification and trial recruitment models using XGBoost and LightGBM — always with bias checks across diverse cohorts. The infrastructure side was just as important: end-to-end ML pipelines on AWS and Azure, from raw ingestion through to serving, running across multiple concurrent projects. I brought generative AI into clinical workflows — document automation, diagnostic support — and set up the MLOps scaffolding (versioning, monitoring, retraining) to keep it reliable. The Tableau and Power BI dashboards I delivered became daily tools for clinical and ops teams tracking enrollment in real time.

This was the role that turned me into a medical AI researcher. Working with clinicians at ILBS, I built sepsis risk prediction models for ACLF patients — XGBoost and Random Forest with careful threshold tuning, so that model outputs actually mapped to clinical decisions (published APASL 2023). On the imaging side, I implemented U-Net and CNN models for brain tumor segmentation from MRI, validated against radiologist-annotated ground truth using dice coefficient and cross-fold validation (IEEE ICAII 2023). Working directly with clinicians forced me to care about clinical meaning, not just statistical metrics.

My entry into ML was through biology. At Centyle Biotech, I built pipelines to make sense of complex genomic and proteomic datasets — clustering, dimensionality reduction, classification — working in Python and R. I collaborated closely with domain scientists to define features and validation workflows that captured actual biological variability, not just data artifacts. It taught me to iterate carefully and document rigorously before anything touched production.