Fedy Ben Hassouna

• Worked on deploying and integrating AI models on a cloud platform (Azure AI Studio) , ensuring they align with the current billing framework and do not require external payment setups. • Compared two models, GPT-4o and Mixtral, to manage the billing framework within Azure, ensuring operational efficiency and cost-effectiveness

Developed a machine learning model to detect defects in wind turbines by analyzing sensor data and performing data preprocessing. The objective of this project is : - to identify early signs of failure, enabling predictive maintenance. - to reduce downtime, lower maintenance costs - to extend turbine lifespan, ensuring optimal performance.

Developed a sentiment analysis tool to determine the sentiment (positive, negative, neutral) of Twitter tweets. Utilizing sentiment analysis libraries like TextBlob, I implemented a solution to: - Accurately analyze tweet data and gauge public opinion. - Enhance understanding of social media trends and user sentiment. - Support data-driven decision-making by providing clear insights into online conversations.

◦ Developeed RAG Agentic services and dashboard: implemented backend with authentication, streaming RAG endpoints, LLM integration, vector search and embeddings, caching and async streaming; built admin dashboard for document ingestion, vector store sync and user management, improving data quality, reducing latency, enabling semantic access to large volumes of business and operational documentation to support internal knowledge management and decision-making. ◦ Built automated EAN-driven eBay.fr scraping pipelines with shop-priority selection, ML-assisted attribute extraction, robust ETL and data cleansing, incremental real-time updates and workflow orchestration to improve data quality, throughput and scalability.

◦ Developed a DSM-5 Knowledge Graph-based diagnosis agent that validates psychiatric diagnoses from patient statements , extract symptoms and generates differential diagnoses with explainable reasoning. Achieved 94% accuracy across 300+ disorders with interactive interface, improving diagnostic consistency by 20%. ◦ Designed and implemented a custom knowledge graph (Statements-Diagnosis-Symptoms) to ground the agent’s reasoning and integrated it into a multi-agent system, reducing researchers’ manual verification workload by 40%.

◦ Led end-to-end development of an AI-driven predictive maintenance platform cutting industrial pump downtime 70% ; engineered LSTM time-series forecasting on multi-sensor data (RMSE 1.4%) and autoencoder anomaly detection via log-mel spectrograms (87% precision, 4.6/5 recall). ◦ Architected RAG-LLM agent delivering 72% more specific maintenance recommendations and 58% higher actionability; evaluated via LLM-as-Judge achieving 96% alignment score and 98% source-grounded responses.

◦ Built an AI-powered analytics agent for retail company , enabling data and business analysts to query KPIs (sales, promotions, inventory) in natural language; reduced manual SQL/reporting effort by 80% . ◦ Delivered 7 interactive analytics modules (promotion impact, stock turnover, basket analysis, forecasting,unsold stock detection) used daily by 20+ analysts and managers, accelerating decision-making

• Developed an autonomous garbage collection robot using an ESP32 microcontroller, integrating YOLOv8-based object detection (fine-tuned on a custom obstacle dataset) via a laptop webcam to identify and collect trash items such as bottles and papers, achieving 90% detection accuracy on common trash types. • Implemented real-time navigation and obstacle avoidance using ultrasonic sensors and DC motors, reducing collision incidents by 95% in dynamic environments. • Practical Workshops: Led interactive workshops to guide students through building and programming their own autonomous robots, focusing on automation, AI, and hardware integration.

◦ Developed a deep learning system using DeepForest to detect cork oak trees in aerial imagery, achieving 84.31% accuracy and 91.49% F1 score on validation data. ◦ Created a custom pipeline for biomass estimation based on detected canopy areas and allometric equations, supporting forest inventory and management. ◦ Annotatedhigh-resolution images using VGG Image Annotator and converted them to DeepForest-compatible formats, enhancing model training efficiency