Sofiane Derraz

Architected and deployed end-to-end LLM pipelines handling daily requests; optimized inference latency through model quantization and batching. Designed multi-model orchestration supporting GPT, Claude, and open-source LLMs; implemented intelligent routing reducing API costs by 35%. Built proprietary Retrieval-Augmented Generation (RAG) system leveraging vector databases (PostgreSQL/pgvector) and semantic search; improved content relevance scores by 28% with prompt engineering. Analyzed content usage patterns and user behavior; identified optimization opportunities increasing feature adoption by 22%. Implemented containerized microservices (Docker + Kubernetes) for seamless model serving and auto-scaling; established ML monitoring (drift detection, latency tracking) and automated retraining pipelines using MLflow. Reduced cloud storage costs by 30% through intelligent data versioning and lifecycle management. Built CI/CD pipelines for model deployment with sub-minute deployment cycles

Delivered predictive analytics and forecasting for ticket stock optimization (18% improvement in inventory turnover), tourist flow forecasting with 92% accuracy enabling dynamic pricing strategies, and customer demand modeling to optimize resource allocation. Developed NLP-driven automation bot, automating 80% of routine tasks; implemented sentiment analysis to inform product roadmap. Built real-time dashboards and competitive intelligence tools to support strategic decisions in a tourism-focused context; migrated deployment to Azure with CI/CD, reducing deployment times to minutes.

Led NLP & text classification initiatives: built end-to-end pipeline for extracting signals from 100k+ reviews; trained fine-tuned BERT model achieving high F1 on multi-label classification; engineered 40+ signals from unstructured text to enable data-driven career guidance for learners. Built robust data collection and ETL pipelines for 500k+ reviews with high data quality; delivered incremental data updates and scalable processing.