Machine Learning Engineer at DRIMCO GmbH

Architected a dockerised federated learning pipeline using LayoutLMv3 (classification) and LayoutLM (extraction), improving accuracy by ~7%. Fine-tuned RoBERTa with layer freezing, improving EM by 30.9% and F1 by 32.4%, with MLflow deployment and experiment tracking. Deployed a multi-model inference stack (LLaMA 7B, RoBERTa) on AWS EC2 via Docker-based model-specific endpoints. Built an agentic RAG pipeline (BM25 + FAISS) with dynamic model routing and multithreaded POST inference for tender QA. Engineered a three-tier extraction engine (rule-based + IR + LLM) for requirement analysis with ~73% production accuracy. Developed a ticket-routing microservice for para-level assignment using hybrid topic classification + IR retrieval. Delivered FastAPI microservices with PostgreSQL ETL pipelines and LangChain-orchestrated GPT workflows with ~1s latency.

Curated and labeled 1k+ dataset using LabelStudio; fine-tuned LayoutLM for key-value retrieval, improving accuracy by 5% and precision by 3%. Developed geometric merging and spatial token alignment for inference and a novel LayoutLMv3 dataset curation approach. Consolidated overlapping topic schemas and rebalanced skewed classes using LLM-based data augmentation. Engineered a DistilBERT training pipeline (stratified split, class weighting) for topic classification, boosting accuracy by 7%.

Research Assistant — Machine Learning for Nanomaterials at École Polytechnique Fédérale de Lausanne (EPFL)

Supported a feasibility study for applying AI to nanocellulose material characterization by defining scope, metrics, and approach. Curated AFM/SEM datasets from 200+ studies and built preprocessing pipelines for resolution handling, alignment, and noise reduction. Increased dataset volume ~5x using data augmentation, Neural Style Transfer, and annotated data creation using CVAT. Implemented YOLOv8 and YOLOv5 for instance segmentation in cellulose nanocrystals (CNCs), achieving ~63% and ~59% accuracy respectively.