Ikrame A

AI Engineer

Liebherr Group • Temps plein

Aug 2025 - Present • 9 mos

Developed an end-to-end AI application for automatic extraction and normalization of equipment data from heavy machinery PDF brochures (Liebherr, Komatsu, CAT, Volvo). Built a multi-tab Streamlit app deployed on Databricks Apps, integrating Azure Document Intelligence for PDF parsing, LLMs for equipment classification, and a sentence-transformers embedding model for similarity matching. Implemented a canonical equipment database using Databricks Delta Tables, with automatic alias detection and deduplication via LLM comparison. Set up MLflow experiment tracking to monitor normalization runs, log parameters, metrics, and artifacts. Designed the full data pipeline from raw PDF to structured Excel output with Standard/Optional/Not Available equipment statuses.

AI Engineer

EDF Energy • Temps plein

Jun 2024 - Jul 2025 • 1 yr 1 mo

Developed a production-ready RAG-based chatbot using Nemotron (NVIDIA) to answer user questions in natural language from internal company documents. Collected, cleaned and structured large volumes of unstructured internal data including PDFs, DOCX files and emails, then integrated them into a RAG pipeline combining document retrieval and response generation. Built a vector database for semantic search and designed efficient chunking and indexing strategies to maximize retrieval accuracy. Designed and developed a user-friendly interface allowing employees to interact with the AI assistant in real time. Ensured the solution was robust, scalable and adapted to the company's internal knowledge base.

AI Engineer

Capital One • Temps plein

Dec 2022 - Apr 2024 • 1 yr 4 mos

Contributed to Talentino, an AI-powered recruitment automation platform. Developed proofs of concept for intelligent document processing, focusing on automatic extraction of structured information from CVs and cover letters. Fine-tuned multiple vision and language models including Paligemma, Moondream, and LayoutLMv3 on a dataset of 9000 annotated CV images to interpret visual language level representations (stars, bars, circles) and return a standardized score from 1 to 5. Managed the full fine-tuning pipeline using RunPod, Lightning.ai, TrainML and Weights & Biases for experiment tracking. Evaluated model performance using LangSmith and ROUGE metrics. Developed a Streamlit application to automate model evaluation, allowing the team to compare multiple HuggingFace models on custom datasets without modifying the code.