Maxence Nevoret

– Contributing to international initiatives comparing methane detection and quantification methods: – Stanford Controlled Release Study: Assessing the accuracy of satellite-based methane detection by comparing remote sensing measurements to controlled ground-truth emissions. – MEDUSA Project: Validation and intercomparison of multiple methane detection and quantification methods. Worked alongside international participants to assess performance, identify limitations, and contribute to improving the current state of the art. Building national methane emission inventories by leveraging satellite observations (KAPSARC Collaboration): – Applied the Wind Rotation model to estimate annual methane emissions from infrastructure by aggregating satellite detections and aligning plumes based on wind direction to enhance signal strength and measurement precision. – Extrapolated asset emissions to generate country-level emission estimates. – Improved the existing methodology and assessed model robustness by conducting a comprehensive sensitivity analysis. Developing a real-time pipeline for detecting methane emissions from geostationary satellite video data: – Designed a dynamic destriping algorithm that leverages the high temporal resolution of geostationary observations to correct sensor artifacts. – Implemented plume delineation and background subtraction using Robust Principal Component Analysis (RPCA) to isolate methane anomalies from complex, variable backgrounds.

Developed an end-to-end pipeline to detect methane emissions from EMIT hyperspectral imagery: – Implemented a data harvester to retrieve raw EMIT images from NASA Earthdata – Processed methane-enhanced images using the matched filter algorithm – Created training, validation, and test datasets for supervised learning – Trained a deep convolutional neural network (CNN) to segment methane plumes from methane-enhanced imagery – Quantified detected emissions using the Integrated Methane Enhancement (IME) method

Product Matching Project: – Developed an algorithm to identify equivalent products with different names/ identifiers for a major European retailer. – Applied textual and semantic similarity techniques on product labels and ingredient lists to detect equivalent products – Built custom training and inference datasets from available product data – Trained BERT models to learn product representations – Created a custom data collator to generate meaningful positive and negative sample pairs, enhancing model performance during training Recipe generation & product recommendation Chatbot – Created a chatbot that generates recipes and connects ingredients to real-time store products for a European retailer. – Developed a product search engine using cosine similarity on fine-tuned BERT embeddings for accurate product retrieval. – Integrated GPT-3 and Langchain to build a conversational agent for recipe generation and product suggestions.