Rami Allouch

Contributed to two EU projects (VPP4ISLANDS, FlexCHESS) as part of my role at LIS, Aix-Marseille Université: VPP4ISLANDS — Virtual Power Plant for islands Real-time optimization backend: Re-architected a legacy, static-data optimizer into an API-driven service that ingests external and internal feeds in parallel (async), runs optimization, and exposes results via FastAPI. Operational dashboards: Delivered React dashboards for KPIs and optimization outcomes used by operators and stakeholders. Auction-based energy trading (blockchain): Built a full-stack solution for node-to-node trading. The backend orchestrates auction workflows via deployed smart contracts and uses live telemetry to compute bids/setpoints and submit them on-chain (Python/FastAPI). The React front end provides operator and node views for auction states, bids, and returned setpoints. FlexCHESS — Hybrid energy storage & flexibility services (ongoing) Frontend lead (React): Leading role-based interfaces, user/asset and chess-node management, dashboards, and workflows, with contributions to backend services (FastAPI). Edge-to-cloud device integration: Built a Python module for the Victron Cerbo GX (energy system monitoring device) to republish telemetry to a cloud-hosted MQTT broker on the device OS, plus a FastAPI service that discovers devices over Modbus TCP and enables real-time read/write telemetry.

Speech Recognition System on Microcontrollers | STM32F407-DISC1 Developed a CNN-based speech recognition system for microcontrollers, implemented on the STM32F407-DISC1 board. The project involved designing, training, optimizing, and deploying a deep learning model capable of classifying spoken commands to control embedded devices. Followed the complete deep learning workflow: Data preparation — audio collection, cleaning, preprocessing, augmentation, class balancing, and MFCC feature extraction using Librosa and Audiomentations Model development — CNN design and hyperparameter optimization with Optuna, Keras Tuner, Scikit-Optimize, and Talos. Model optimization/compression — applied quantization, pruning, clustering and other techniques (via TensorFlow Lite) to reduce model size from 15 MB to 120 KB while maintaining 84% test accuracy Deployment — integrated and validated the optimized model using STM32CubeIDE and X-CUBE-AI, achieving efficient on-device inference within the STM32’s 1 MB flash and 192 KB RAM constraints This project demonstrates the feasibility of embedded AI speech recognition on resource-limited microcontrollers.

Worked on a project in collaboration with psychiatrists from the Regional Hospital of Sfax, focusing on factors influencing mental health problems during the COVID-19 pandemic. Developed a machine learning model to detect mental health issues such as depression and anxiety. Applied interpretable machine learning techniques to identify the most influential features associated with each condition using global interpretability methods such as Partial Dependence Plots (PDP) and Permutation Feature Importance. Additionally, provided explainable predictions for individual patients—indicating the presence or absence of a disorder and the underlying reasons—using local interpretability techniques like LIME and SHAP.

Developed a full-stack web application that enables employees to log their working days, completed tasks, and justified absences. The platform includes an admin dashboard for managing employees, missions, and overall activity tracking. Tech stack: React.js, Spring Boot, MySQL