Youcef Benkhedda

Spearheaded research and development of Generative AI and Foundation Model solutions for healthcare applications. Designed privacy-preserving synthetic data pipelines for NHS, enabling research within strict data compliance. Implemented training strategies and fine-tuned foundation models (Llama, Gemma, Qwen) using PEFT/LoRA and 4-bit quantization (AWQ) to optimize inference on resource-constrained hardware. Architected Agentic Workflows with LangGraph and integrated RLHF to align outputs with expert guidelines. Built a validation framework for generative outputs including Membership Inference Attacks (MIA) and semantic checks using ClinicalBERT embeddings. Established reproducible ML pipelines with MLflow.

Led technical architecture for the £3.6M 'Our Heritage, Our Stories' project, transforming unstructured archival data into structured Knowledge Graphs using Representation Learning and large-scale extraction. Designed a zero-shot Relation Extraction pipeline via Natural Language Inference (NLI). Developed and deployed Neural Machine Translation pipelines (AraT5v2, mBART) with Joint Regional training. Built scalable data processing workflows handling 1M+ records to populate high-dimensional Knowledge Graphs (Neo4j). Authored and contributed to peer-reviewed work on Representation Learning and Hate Speech Detection (benchmarks like AraTar).

Developed high-throughput inference systems and stance detection models for real-time global sentiment analysis. Implemented a Network-Based classification approach (graph topology) to model user stance, achieving F1-score of 0.95. Deployed GPU-accelerated NLP pipelines (Stanford Stanza) for processing 5.5M tweets daily. Built a full-stack analytics platform using Dash/Plotly to visualize high-dimensional ML insights for stakeholders.