Lachouak Heni

AI & Automation Engineer chez Nehos Groupe depuis août 2025. Je conçois des agents LLM et systèmes RAG en production pour l'automatisation marketing (pipelines n8n de publication et génération SEO), le service client (agent vocal multi agent avec LangGraph, Twilio et RAG/KAG) et la legaltech (extraction OCR et fine tuning Qwen 20B via QLoRA sur documents notariaux). Stack principale : LangChain, LangSmith, RAG/KAG, n8n, microservices.

I completed a rewarding internship at Autolive, where I successfully led several artificial intelligence projects that combined innovation with technical rigor. 🔧 Projects Completed: **Steering Wheel Defect Detection System: I created a proprietary dataset by capturing and annotating approximately 3,000 images of real defective steering wheels. Utilizing the Roboflow platform, I developed a high-quality dataset featuring two defect classes: adhesive marks and sewing errors. With this dataset, I fine-tuned the YOLOv8 model, achieving a detection rate of 98%. The model was subsequently deployed on an NVIDIA Jetson board within a system that I developed entirely. This system includes a camera, an ultraviolet LED, a display, and a software interface designed using Django and Flask to ensure optimal automation of quality control. **AI-Powered Reconstruction and Semantic Analysis of Quality Report Archives. : Developed an AI pipeline using OCR and NLP techniques to reconstruct quality report archives. Performed semantic analysis to extract defect and performance rate.

Deep Learning System for Glaucoma Screening in Fundus Images As part of my research in artificial intelligence, I developed a deep learning-based glaucoma screening system that integrates advanced machine learning techniques to enhance diagnostic accuracy and efficiency. Project Objective: The aim of this project was to create an automated solution for glaucoma screening using fundus images, enabling healthcare professionals to detect this ocular disease early, quickly, and reliably. Given that glaucoma is a leading cause of irreversible blindness, the goal was to provide a tool that reduces reliance on human expertise while improving diagnostic precision. Technical Approach: I designed a hybrid architecture that combines a CNN-Swin Transformer model, leveraging the spatial feature detection capabilities of Convolutional Neural Networks (CNN) with the global relationship capturing power of transformers. This hybrid approach optimized the classification of fundus images to differentiate glaucoma cases from healthy ones. Data Preparation: I worked with the JustRAIGS dataset, applying preprocessing techniques such as image enhancement via CLAHE (Contrast Limited Adaptive Histogram Equalization) and data augmentation to enhance the model's robustness. Optimization and Results: The model was fine-tuned through hyperparameter optimization, blending mathematical theory with practical experience. By adding additional CNN layers, I achieved an impressive 99% accuracy across most performance metrics, a significant result in the realm of AI-assisted medical screening. Academic Recognition: This project culminated in the publication of the paper titled "Hybrid CNN-Swin Transformer for Glaucoma Screening in Fundus Images," which was accepted as a Full Paper with Oral Presentation at the AICCSA 2024 conference in Sousse, Tunisia, highlighting the importance of this solution in automated ocular disease screening.