Hi, I'm Mohamed Amine AROUS

Developing a Neuro-Symbolic AI pipeline fusing Foundation Models with an enriched Knowledge Graph to improve semantic scene understanding for autonomous warehouse robots.

• Designed and implemented a Knowledge Graph enrichment pipeline achieving 89% enrichment accuracy, improving the reliability of injected semantic knowledge.
• Integrating structured knowledge sources (Wikidata, domain ontologies) to ground robot perception and reasoning.
• Leveraging environment plans and business rules to guide planning in warehouse navigation scenarios.

Developed CosyVoice2-EU, an open-source adaptation of a state-of-the-art Transformer-based voice cloning model for European languages, achieving performance exceeding strong open-source baselines while remaining competitive with a leading commercial speech synthesis system (ElevenLabs).

Publications:

• EUSIPCO 2026 – Efficient adaptation method of a SOTA Voice Cloning model to European languages

• Built a serverless RAG-powered assistant widget for my portfolio site, using FastAPI on AWS Lambda with hybrid BM25/embedding retrieval to answer visitor questions grounded in my own career data.
• Engineered for near-zero operating cost and validated with 413 passing tests, achieving zero hallucinations across 18 evaluation questions.

• Developed machine learning models in Python (Scikit-learn, PyTorch, Pandas, NumPy) to evaluate similarity between microelectronic components.
• The project delivered a custom AI model that reached 98% accuracy, surpassing the performance of the company's internal solution.
• Collaborated within a multidisciplinary engineering team to design and validate the final solution.

• Designed and deployed a fully automated CI/CD pipeline on AWS to take a web app from Git commit to production with zero manual steps.
• Automated artifact management, build, deployment, and release orchestration using CodePipeline, CodeBuild, CodeDeploy, and CodeArtifact.
• Deployed production workloads on EC2 with IAM-secured, git-triggered releases on every GitHub push.

• Developed a local Retrieval-Augmented Generation (RAG) backend using FastAPI and Ollama, implementing a modular architecture integrating document ingestion, embedding generation, semantic retrieval, and LLM-based response generation.
• Designed REST API endpoints separating indexing, retrieval, and inference services, enabling efficient local deployment using open-weight language models.

• Developed a deep understanding of state-of-the-art Vision-Language-Action models.
• Explored and tested different approaches to improve CoIN and AIUTA.
• Benchmarked AIUTA with different LLMs on the school's GPU cluster.

• Built a full-stack data pipeline to identify exoplanets from light curve data, achieving ~94.3% accuracy.
• Trained multiple models (CNN, Autoencoder, MLP).

• Developed AI agents for candidate body-language analysis during video interviews and automated resume SWOT evaluation.

• Integrate a dataset via representations from a pre-trained LLM and train classical models on these embeddings.

• Deputy Technical Manager
• Développement d'un prototype d'une ruche intelligente pour aider les apiculteurs à améliorer leur production et la qualité du miel.

• Co-auteur d'un rapport scientifique sur les méthodes d'apprentissage automatique et les approches distribuées.

• Stage d'un mois pour une immersion dans le milieu professionnel des infrastructures de télécommunications.

Weekly tutoring in pairs with high school students in science and mathematics.

• Weekly support aimed at strengthening knowledge and promoting equal opportunities.
• Reinforcement of basic scientific concepts and preparation for post-baccalaureate courses.

• Organization of DSA courses to strengthen problem-solving skills.
• Promotion of national competitions and events.

• Improved global rank in IEEEXtreme from 1370 (17.0) to 491 (18.0) with team "DebuggingTrio".