Félix MORTAS

📞 Assessing the Carbon Footprint of Virtual Meetings: A Quantitative Analysis of Camera Usage

Abstract: This paper quantifies the carbon emissions related to data consumption during video calls, focusing on the impact of having the camera on versus off. The findings regarding the environmental benefits achieved by turning off cameras during meetings challenge the claims of some prevalent articles. The experiment was carried out using a 4G connection via a cell phone to measure the varying data transfer associated with videos. The outcomes indicate that turning the camera off can halve data consumption and associated carbon emissions, particularly on mobile networks. The paper concludes with recommendations to optimize data usage and reduce the environmental impact during calls.

Comments: [v1] 4 pages, 3 figures, submited and accepted as a short paper by IARIA GREEN 2025 with some fixable issues that must be addressed before the article is ready for publication (Scientific and technical; English & punctuation; Sections and presentation flow)
[v2] Corrected with the help of checkmymanuscript.com, and rearrangement of sections and presentation flow

Read the article

🚀 Intelligent membership automation (Association Toits Vivants)

As treasurer of the association, I transformed a manual and time-consuming administrative process into a 100% autonomous workflow combining serverless computing, AI, and no-code technology.

The concept: A pipeline that intercepts HelloAsso payments, updates Google Sheets records, and ensures personalized member onboarding.

Technical highlights:

• Robust AI parsing and judge: Use of the Mistral API (LLM) to extract data from complex emails. Unlike traditional Regex, AI guarantees total resilience to layout changes and user data entry errors.
• Optimized serverless architecture: Business logic developed in Python (Object Oriented Programming) on AWS Lambda. I favored lightweight libraries (requests) to minimize cold starts and maximize responsiveness.
• Orchestration & security: Flow controlled by Make.com with secure communication via tokens (environment variables) and two-way validation.
• Software quality: Code structured in decoupled services, fully covered by unit and integration tests for easier maintenance.
• Costs: 100% free. Also, small LLMs have been choosen to minimize API costs and environmental impact on scale.

Results: An instant registration process, an error-free database, and a streamlined member experience (email & WhatsApp), freeing the association from administrative constraints.

View the project on GitHub

🌍 Generate L10n – VSCode AI extension for Flutter internationalization

Development of an innovative VSCode extension designed to eliminate the hassle of internationalization (L10n) in Flutter projects. It automates string detection and translation by combining the native user experience of VSCode with the power of LLMs.

The concept: An integrated visual interface that allows you to select Dart files and use AI to instantly generate .arb files and update the source code, followed by automatic execution of Flutter commands.

Technical highlights:

• Multi-provider AI & Orchestration: Integration of LangGraph to orchestrate complex file modification logic. Compatible with cutting-edge models (Mistral, Gemini, OpenAI) for maximum contextual accuracy.
• Native VSCode UI: Full implementation of an interactive Tree View in TypeScript in the Activity Bar, offering granular management of files to be processed.
• End-to-end workflow: The extension doesn't just translate; it intelligently modifies the Dart code, manages security backups, and triggers the Flutter code generator (gen-l10n) in the background.
• Robust architecture: Automatic detection of project configuration (pubspec.yaml) and secure management of API keys via VSCode settings.

Result: Significant time savings for Flutter mobile developers, transforming a manual task that took several hours into an AI-assisted process that takes just a few seconds, while ensuring translation consistency and the ability to easily revert changes.

🤖 Autonomous Agentic System – GAIA Benchmark Solver

Final project for the Hugging Face Agents Course. I developed a high-level autonomous agent capable of solving complex, multi-step tasks from the GAIA Benchmark (General AI Assistants), involving real-world tool usage and multimodal reasoning.

The concept: A robust agentic workflow built with LangGraph that follows a Thought-Action-Observation cycle to decompose 20 validation queries into executable steps, navigating through technical constraints like API rate limits and data extraction challenges.

Technical highlights:

• Resilient Model Orchestration: Implemented a fallback & routing strategy using Gemini 2.5 Pro as the primary brain, with automatic switching to Gemini Flash, Mistral, or Groq-hosted models to bypass free-tier rate limits without interrupting the execution flow.
• Advanced Tool Engineering: Instead of overloading the context window with many small tools, I developed a utils.py library of complex functions. The agent uses a refined set of "Super-Tools" (Web Search, Excel manipulation, Audio Transcription, API interaction) that handle internal logic complexity autonomously.
• Multimodal Innovation: Engineered a custom Video Analysis sub-agent. Since no free direct video-to-text API was available, I built a pipeline that intelligently extracts frames and metadata to reconstruct temporal context for the LLM.
• Custom RAG Architecture: Integrated ChromaDB with a specialized retrieval algorithm optimized for the specific nuances of the GAIA dataset, ensuring the agent retrieves only the most relevant context for its reasoning steps.
• Observability & Evaluation: Self-hosted LangFuse locally to monitor traces, evaluate agent costs, and debug the Reasoning-on-Action (Re-Act) loops without incurring cloud platform fees.
• Full-Stack Deployment: Interface built with Gradio and hosted on Hugging Face Spaces, managed via Git for version control and CI/CD.

Results: Successfully validated 16 "Level 1" GAIA tasks, demonstrating a high degree of autonomy in tool selection and the ability to maintain long-term state across multiple reasoning cycles.

🧠 Flasholator – Spaced Repetition Flashcards & Instant Translation

Flasholator is a mobile application designed to bridge the gap between instant translation and long-term vocabulary retention. It allows users to translate unknown words on the fly and instantly convert them into a personalized study deck.

The concept: A seamless workflow where DeepL handles the translation, and a custom Spaced Repetition System (SRS) ensures you never forget what you've learned, optimizing your study time by focusing on your weakest words.

Technical highlights:

• Optimized UI/UX: Carefully crafted user interface and experience to ensure intuitive navigation and a delightful learning journey.
• Intelligent Learning Algorithm: Implementation of the SuperMemo-2 (SM-2) algorithm in Dart. It dynamically calculates optimal review intervals based on user-reported difficulty and repetition history to maximize memory retention.
• Deep API Integrations: Leveraging the DeepL API for high-accuracy linguistic translations.
• Robust Offline-First Architecture: Built using MVVM with a local-first approach. It utilizes SQLite (Drift) for high-performance offline access.
• Authentication & Authorization: Secure user authentication and authorization using Firebase Authentication and Firestore Security Rules.
• Monetization & Privacy Integration: Production-ready setup featuring RevenueCat for multi-platform subscriptions, Google AdMob for ads, and full GDPR/UMP compliance for user data privacy.
• Automated Localization: Leveraging my custom-built "Generate L10n" VSCode extension to manage multilingual support (French, English, Spanish) via LLM-generated ARB files.
• Modern CI/CD Pipeline: Automated build and distribution flows using GitHub Actions for Google Play Store and Codemagic for iOS App Store Connect.

Result: A scalable, production-grade Flutter application that provides a friction-less learning experience, turning a simple translation tool into a powerful personal tutor.

🎓 Data Science Graduate Program – End-to-End ML & MLOps Specialization

An 18-month apprenticeship at AXA with CentralSupélec, focused on solving complex business problems through data engineering, machine learning, and cloud-native deployment.

The concept: From raw data exploration to production-ready AI, I completed 9 industrial-grade projects covering the entire Data Science lifecycle, including deep learning, big data, and MLOps.

Technical highlights:

• Machine Learning & Deep Learning: Expertise in supervised learning (classification/regression) and unsupervised learning (clustering). Advanced mastery of unstructured data via NLP and Computer Vision.
• MLOps & Cloud: Industrialization of models via Cloud (AWS/Heroku) architectures. Implementation of CI/CD pipelines, lifecycle management with MLFlow, and creation of scoring APIs (Flask) with Data Drift monitoring.
• Big Data & Scalability: Manipulation of large volumes of data with SQL and distributed processing via PySpark on AWS EMR cloud clusters.
• Business Intelligence & Reporting: Designing interactive dashboards (Streamlit) and simplifying complex results for decision-makers. Using explainable AI (SHAP) to justify each prediction.
• Scoping & Ethics: Managing projects using Agile methodology, drafting methodological notes, and strictly integrating GDPR compliance from the design stage.

Results: A comprehensive portfolio of production-ready solutions, demonstrating the ability to transform abstract business needs into scalable, ethical, and high-performing AI products.

View full portfolio on GitHub