Paramita Choudhury

I am a computational scientist based in Dresden, Germany, working at the meeting point of high-performance computing, machine learning, and formal reasoning.

I recently completed an M.Sc. in Computational Modeling and Simulation at TU Dresden (track: Logical Modelling). Before returning to academia, I spent six and a half years as a software developer and QA engineer at Cognizant, working inside large-scale Oracle production systems for global insurance clients across Asia and the USA — terabyte-scale data where correctness was non-negotiable.

Most recently, as a research assistant at ScaDS.AI (TU Dresden), I scaled a transformer-based genomic classifier (DNABERT-2) from a single GPU to eight using PyTorch DDP and NCCL on an HPC cluster, and taught myself performance tracing with Score-P and Vampir to understand where the time really goes in multi-GPU training.

Research interests

Performance and scalability analysis · GPU and parallel programming · HPC performance tools and tracing (Score-P / Vampir) · machine learning for scientific computing · numerical methods for PDEs · knowledge representation and reasoning

Selected work

Score-P / Vampir performance analysis of multi-GPU DDP training — the finding (blog) · debugging field guide (dev.to)
Master’s thesis — Explaining Projected Answer Sets Using Faceted Reasoning — PDF · code
PDE solver & data-driven error corrector for Burgers’ equation — code
Research project — Generalising Stable Sets of Cooperative Games Using Abstract Argumentation — report