AMINE HAMOUCHI

Arabic text varies across domains and writing styles. Comparing representations and model families requires a clear preprocessing and evaluation workflow.

• Arabic normalization, cleaning and separate tweet/review preprocessing paths.
• Experiments with sparse text features, embeddings and feature selection.
• Classical ML, neural-network and AraBERT comparison notebooks and scripts.
• Saved results report AraBERT macro F1 of 0.815 on ASTD binary classification and 0.800 on the review multiclass task.
• Scores are historical outputs from the saved comparison CSV, not newly reproduced benchmarks. The report provides academic context; contributor credit belongs to the project team.

A trained model is only one part of an ML application. Experiments need tracking, deployment needs a quality gate, and the running service needs visibility.

• SciBERT text-classification training with MLflow experiment tracking.
• Jenkins workflows and helper scripts for training, evaluation and deployment checks.
• FastAPI serving and Prometheus/Grafana monitoring configuration.
• Data-drift and monitoring scripts, with existing data and utility tests.
• Adapted from Made With ML, with team contributions. The saved run reports overall F1 of 0.428 on 191 samples. The project emphasizes lifecycle engineering; quality-gate thresholds are configurable.

A sensor-driven application needs a usable connection workflow as well as a place to display and retain incoming activity data.

• Bluetooth device discovery, connection state and characteristic subscription.
• Stream-based step-count updates in a Flutter interface.
• Local SQLite storage for activity history.
• Enclosure design assets supporting the wearable prototype.
• This case study covers the mobile BLE interface, local storage and enclosure design. Device firmware and step-detection accuracy require separate hardware validation.

Multiple web applications need a common foundation for identity, navigation, shared UI and backend data access.

• A pnpm/Turborepo monorepo with five Next.js application directories.
• Shared packages and common platform components.
• Central authentication, role-based access and PostgreSQL integration.
• NestJS/Prisma backend code; a separate Spring Boot backend also exists in the project workspace.
• An internship/team platform with some applications still in development. The overview describes the architecture, not a claim that every portal is production-complete.

Environmental measurements provide a practical setting for preparing tabular data and evaluating regression models.

• A Python/scikit-learn regression study using Random Forest.
• Prediction targets include CO and NO₂.
• The academic report documents the analysis and evaluation context.
• Presented as an academic report-based study. The report is available below; a reproducible source-code release is not currently provided.

Invoicing and warehouse inventory that keeps working without a connection.

• Flutter app for invoices and stock.
• Offline-first sync.
• Supabase and Docker behind it.