I am a Machine Learning Researcher at NuMind, where I bridge the gap between cutting-edge NLP research and scalable production systems. My work focuses on the full lifecycle of Large Language Models, from dataset curation and training to quantization and deployment.

At NuMind, I specialize in training and fine-tuning multimodal LLMs for complex information extraction tasks. This includes:

• Structured Extraction: Training models to populate user-provided JSON templates with high precision.
• Unstructured Extraction: Transforming diverse documents into clean, functional Markdown.
• Production & Optimization: Leading model quantization efforts and managing production environments, primarily leveraging vLLM.

I hold a PhD in Machine Learning from Dublin City University, where my research focused on Natural Language Processing and the Transformer architecture. In my thesis Breaking Language Barriers: Reimagining Machine Translation as Style Transfer, I explored reimagining machine translation as a style-transfer problem (CycleGN, developed novel metrics for evaluating NMT-generated text (Tokengram_F, Embed_Llama).

Before transitioning into Neural Networks and Machine Learning, I earned two Master’s degrees in Cybersecurity, providing me with a robust foundation in systems analysis and engineering.

• Core: Python, PyTorch, Transformers, PEFT, vLLM, llm-compressor.
• Specializations: Multimodal LLMs, Fine-tuning, Quantization, Information Extraction, Dataset Engineering.
• Previous Research: Neural Machine Translation, Style Transfer, Lossless Text Compression, Encoder-Decoder transformers.

• LinkedIn: linkedin.com/in/sorendreano