Sergio Winter

- Recruitment of Data, ML and MLOps skills for Toulouse agency - Referent for training and support of new arrivals - Identification and development of business opportunities

- Development of a conversational agent solution for M6 collaborators - Design and implementation of a GenAI platform (Bedrock, LiteLLM, ..) - LLM agent and LLMOps monitoring put into production in 4 months (Langfuse) - Development of an automated response quality assessment system (LLM as Judge) - Securing against prompt injection

- Designed the MLOps architecture for the FlywithAI project. - Created a team composed of Data Scientists, Data Engineers, and DevOps profiles.

- Comparative study of LLMOps solutions.

Customer Scoring System : - Optimisation: redesign of the code allowing a parallelization of the processing tasks and a reduction of the memory consumption. This allowed to go from 12 hours of execution to less than 3 hours. - Migration of a scoring application: writing of a SageMaker pipeline for pre-processing, hyperparameter research, training and inference (projection on the next years and estimation of the customer value).

Logistic optimization problem - Dataiku on AWS - Development of forecasting models (timeseries) to better anticipate the flow of maritime containers around the world (Prophet / Regression Models) - Development of algorithms and pre-processing codes (Dataiku) - Deportation of heavy optimization models (Operations research / EKS cluster/ AWS Batch)

- Development of a damage detection system on maritime containers, thanks to the analysis of photos taken with a smartphone application - Training of state of the arts models on the SageMaker platform - Implementation of a damage annotation system (SageMaker Ground Truth) by business experts, allowing the creation of training datasets, as well as a quality control of predictions - Self-Supervised & Active learning pipelines to reduce annotation requirements (-94%) - Setting up a vector database (FAISS) to store image embeddings

Kubeflow on AWS: - Implementation of a MLOps platform with Kubeflow (EKS). - Development of satellite image processing demonstrations to showcase functionalities and facilitate data manipulation by scientists. - Optimization of the scalability and performance of all the usecase processing, which allowed to increase its scope of action. The solution allows the processing of large volumes of data (satellite images), for a very competitive cost.

Smart camera / Intrusion detection system : - Development of a POC for physical intrusion detection, thanks to the analysis of video streams from video protection cameras (OpenCV, YOLOv3) - Creation of different prediction models on the SageMaker platform - Optimization of the prediction speed, to allow an almost instantaneous detection of intrusions, in order to alert the user as soon as possible

- Development of a prediction model to detect fraudulent traffic - Consolidation of the ML methodology - Deployment of the model on SageMaker - Automation of model training (XGBoost, CatBoost, Sklearn) - Training of the Data Scientist team (MLFlow, SageMaker, AWS) - Extraction of patterns to facilitate the understanding of predictions by the business - Synthetic structured data generation

- Building a recommendation engine for media content (collaborative and content-based). State of the art research and implementation of the algorithm.