Monseej Purkayastha

Forschung zu zuverlässigen Entscheidungen unter Unsicherheit: Off-Policy-Evaluation mit partieller Identifikation und Analyse von Fehlern beim logischen Schließen großer Sprachmodelle (LLMs). • Designing and running experiments on partial-identification methods for off-policy evaluation without support overlap, encoding expert knowledge as linear constraints on policy value. • Evaluating and characterizing LLM reasoning failure modes through an imprecise-probability framework for uncertainty modelling.

Masterarbeit: multimodales Diffusionsmodell für stilisierte 3D-Menschenbewegungen aus Text und Referenzvideo; Erstautor beim CVPR-2026-Workshop MOMA. • Diffusion model conditioned on CLIP text and ViViT video features, disentangling motion content from visual style. • 77.8% vs 68.1% style-recognition accuracy against SMooDi (ECCV'24), 2× lower FID, 70% zero-shot style accuracy. • Multi-view synthetic-data pipeline (SMPL, Blender/pyrender). • Paper: VISTA, MOMA Workshop at CVPR 2026.

Angewandtes ML auf verrauschten Prüfdaten (zerstörungsfreie Prüfung) und Steuerungssoftware für ein mobiles Ultraschall-Messsystem im Kundeneinsatz. • Rail4Future: ML pipelines estimating rail stress from noisy NDT sensor data; -42% out-of-distribution error on new sites. • Found environmental domain shift to be the main failure mode; mitigated via augmentation and robust feature selection. • UER IV: Dockerised Flask app to control and monitor a portable ultrasonic measurement system at client sites.

Tutorien für über 50 Studierende in Statistischer Sprachverarbeitung (NLP) und Neuronalen Netzen. • Led tutorials for 50+ students across 2 courses: Statistical NLP and Neural Networks: Theory and Implementation (Prof. Dietrich Klakow). • Designed PyTorch and NLP programming assignments and problem sets on neural-network theory. • Supported students through tutorials, assignments and exam preparation.

Entwicklung und Betrieb großvolumiger ETL-Datenpipelines für über 9.000 Firmenkunden sowie Migration nach Azure. • High-volume ETL pipelines for 9,000+ enterprise clients (IBM DataStage, Oracle DB, Unix, SAS DI Studio). • Automated testing, monitoring and data validation in CI/CD: -35% manual effort, +12% test coverage. • Incident triage and root-cause analysis for business-critical migrations. • Migrated Oracle ETL workflows to Azure Data Factory and Azure SQL, with CI/CD via Azure DevOps.

Note: 1,4 (deutsche Skala). Schwerpunkte: Maschinelles Lernen, Computer Vision, 3D-Computer-Vision, Robustheit von ML-Modellen, effizientes Training von LLMs. Grade: 1.4 (German scale; 1.0 = best). Coursework: Machine Learning, High-Level Computer Vision, 3D Computer Vision, Robustness in ML, Efficient Training of LLMs.

Schwerpunkte: digitale Signal- und Bildverarbeitung, Algorithmen und Datenstrukturen, Regelungstechnik, Datenübertragung und Netzwerke. GPA: 7.56/10. Coursework: Digital Signal Processing, Digital Image Processing, Data Structures & Algorithms, Control Systems, Data Communication and Networks, Signals and Systems.