Mohamed Abi

Automated metadata extraction from long technical reports using self-hosted LLMs on GPU clusters; designed an end-to-end evaluation framework: Level-1 entity detection and Level-2 attribute extraction, precision / recall / per-attribute F1, document-complexity analysis — macro-F1≈0.94; benchmarked open-weight models (Mistral, Llama, Gemma) through vLLM (L40S/H100 GPUs), ClearML tracking; Pydantic structured extraction and ontology-based validation against hallucinations.

RAG + LLM pipeline (semantic chunking, embeddings, hybrid retrieval) generating reader-tailored summaries of risk reports; FastAPI API in production. Co-author of a peer-reviewed INFORSID 2026 publication on automatic LLM-based risk-report generation.