Hamza Gafsi

- As part of the Innovation team within the Presales department, I design and develop innovative AI projects and use cases for telecom operators worldwide such as: Anomaly Detection & AI-Based Customer Experience Management. - Designed the AI logic behind a next-gen CEM / CX scoring solution, including the CEM Investigator Agent - an agentic system (RAG over a Neo4j network graph, live telemetry) that autonomously diagnoses subscriber QoE issues, cutting investigation from hours to minutes. - Built an E2E AI sentiment analysis product for telecom operators: LLM-based scoring, topic classification, and competitor benchmarking over complaints from 6 sources, with auto-generated reports. Deployed across 4 operators. - Built a customer segmentation model from mobile-app usage behavior, surfacing 4-6 actionable segments for marketing targeting. - Collaborated with customers and supported the sales team by presenting and demonstrating AI solutions. * Tools: Agentic AI, LangGraph, RAG, Neo4j, pgvector, Vertex AI, Langfuse, LangSmith, Ragas, MLflow, Airflow, Docker, Kubernetes, AWS, Vertica, Spark, Superset, Grafana, Streamlit

- Developed and deployed GNN models to optimize business process efficiency. - Enhanced data pipelines using Distributed Computing Frameworks to improve data quality, scalability, and robustness, reducing processing time by ~50%.

- Designed, built, and deployed a real-time analytics platform by collecting and processing large-scale security logs. - Created interactive dashboards to visualize and monitor system activity.

- Built a Voice & SMS fraud detection model achieving ~89% precision. - Developed an NPS prediction model (AUC ~0.85) enabling churn-risk targeting for customer experience teams.

- Built a RT energy disaggregation platform using Linky data, processing streaming consumption data via Kafka and Spark achieving ~91% per-appliance load identification accuracy.

- Designed and implemented telecom fraud detection solutions using Big Data and Machine Learning. - Developed predictive models to uncover suspicious usage patterns, reducing fraud risk and improving network security.