A sentiment analysis model for Moroccan Darija (a dialect with limited NLP resources) classifying text into positive, negative, or neutral categories. The pipeline covers data collection & integration, preprocessing, embedding generation, and fine-tuning of the pretrained DarijaBERT model using LoRA (Low-Rank Adaptation) for parameter-efficient adaptation without full retraining, built with Hugging Face Transformers and PyTorch.
Interactive Gradio chatbot combining LLaMA 3.1 and Neo4j knowledge graphs with RAG for accurate natural-language answers.
Offline agent that turns plain French, English, or Arabic questions into verified SQL over a 117K-row Caterpillar support-claims database, validated at 87% execution accuracy on a 100-question, 3-language benchmark.