Mohamed Yahdhih Sidi

E-commerce product classification pipeline (12M+ products): hybrid approach using vanilla models + zero-shot LLM, accuracy from 60% to 90% across 4 depth levels. Optimization of production RAG pipelines: hybrid search (dense + sparse retrieval), reranking, integration of heterogeneous data sources (APIs, relational databases, unstructured documents). Multi-agent orchestration with LangGraph: asynchronous workflows, complex task chaining, structured knowledge base management. LLM monitoring with Langfuse: tracking costs, latencies, output quality, degradation detection, and alerting. End-to-end industrialization: Docker, Kubernetes (Azure AKS), CI/CD, Terraform, 40% reduction in inference costs.

Technical leadership on LLM production deployment: evaluation, fine-tuning, transitioning proprietary models to open-source (Mistral, LLaMA) with systematic evaluation frameworks. Automated evaluation pipelines: benchmarking, quality metrics, regression testing, and continuous monitoring.

End-to-end automated ML pipelines: training, validation, deployment, and production monitoring. Large-scale data processing with Python, Elasticsearch, and PySpark, CI/CD for ML workflows.