Samir-Guenchi/Ramy-Sentiment-Intelligence

Ramy Sentiment Intelligence: AI-powered Product Review Analysis for the Algerian Market. # 🟠 Ramy Sentiment Intelligence > **Enterprise-grade Arabic sentiment analysis fine-tuned on real Ramy customer voice** > Built for **AI EXPO 2026** · Arabic · Darija · French · NVIDIA H100 --- ## 📌 Overview **Ramy Sentiment Intelligence** is a production-ready NLP system that classifies Arabic product reviews into **positive**, **negative**, and **neutral** sentiment. It was designed specifically for Algerian Arabic (Darija) and code-switched French/Arabic text — the real language customers use when talking about Ramy beverages. The system is served through a live enterprise dashboard featuring real-time inference, model insights, and data exploration. --- # IMPORTANT - MODEL FILE NOT UPLOADED TO GITHUB ## READ THIS BEFORE RUNNING INFERENCE The trained model weights are intentionally **not uploaded** to the repository (GitHub file size limits). To use prediction/inference, you must do the following: 1. Run the training notebook completely: - `h100-finetune-sentiment-fixed.ipynb` 2. Get the exported zip file: - `ramy_h100_finetuned_model.zip` 3. Extract the zip. 4. Use the extracted model folder for inference (pipeline/model path). 5. If using the web app, set: - `FINETUNED_MODEL_PATH= ` Without this step, model endpoints will not run correctly. --- ## 🚨 The Problem Sentiment analysis in North Africa is unsolved at the dialect level: - Reviews mix **Modern Standard Arabic**, **Algerian Darija**, and **French** in a single sentence - Off-the-shelf multilingual models fail on code-switched Maghrebi text - Labeled data in this domain is extremely scarce (~1,500 samples) - Most pipelines suffer from **evaluation leakage** - inflated scores that don't hold in production This project solves all four challenges simultaneously. --- ## 🧠 Model & Architecture | Component | Details | |-----------|---------| | Base model | `aubmindlab/bert-base-arabertv02` (AraBERT v2) | | Task | Sequence classification (3-class sentiment) | | Max sequence length | 256 token …