Leveraging cutting-edge technologies and methodologies to create innovative solutions for clientsEngaged in advanced NLP tasks, primarily focusing on extending LLaMA and Llama 2 models through enhanced pre-training and fine-tuning. Leveraged state-of-the-art PEFT methods, including Q-LoRA, to improve instruction following via RLHF. Additionally, fine-tuned sentence-transformer models for specialized text classification purposes.I have led the development of a unique software solution utilizing the renowned AI image generation model Stable Diffusion. This initiative harnesses collective intelligence and user preferences, introducing a novel way to interact with the model. My role encompassed all aspects of the software's functionality, with a focus on Python for model operation, as well as the creation of a robust backend and database infrastructure for the application.I've used blockchain and smart contracts to tackle a range of challenges in the Web3 space, which included applying decentralized finance techniques to a new area. This effort contributed to the development of a unique solution. In this project, I worked closely with different teams to streamline processes, achieve our goals, and stay updated with the rapidly changing Web3 landscape.

Providing an API for content moderation for some of the biggest online marketplaces in the world, including Leboncoin (more than 5M messages moderated everyday)Designed and delivered in-depth NLP training programs for engineers, covering applications, theoretical foundations, and practical implementations of SotA models including GPT-3Environment :AWSDockerKubernetes (with Helm and Terraform)TravisSpinnakerJiraDataDog

Creating and conducting Deep Learning training programs for colleagues and clients (180+ students & 20H+ of lectures)• Focused on both theoretical explanations and practical implementations of State-of-the-Art NLP models for English and French languages (special focus on CamemBERT, FlauBERT, GPT-2, ALBERT and ELECTRA)Carried out several Data Science projects for diverse clients• Segmented visitors and tailored recommendations of exhibitors for one of the biggest furniture trade show in Europe (Maison & Object by SAFI)• Implemented a novel customer segmentation methodology emphasizing on efficiency of means of communication for SoloInvest• Undertook initial analyses of the transaction databases of Century21 in cooperation with other colleaguesUse of : PyTorch, scikit-learn, SQL, GCP, AWS

Designed and developed the AI engine for a startup specialized in customer journey analysisDec. 2019 - Jan. 2020• Adapted current state-of-the-art attention model for French language CamemBERT (nov. 2019) based on RoBERTa's architecture (jul. 2019, Facebook AI Research) in order to perform text classifications with very few training examplesDeveloped several automated cryptocurrencies-trading bots for a group of investorsOct. 2017 - Oct. 2019• Devised a prototype of a state-of-the-art algorithm for automated event-driven trading using CNNs applied to neural tensor networks’ event embeddings• Conceived, implemented and automated several innovative strategies to exploit the markets’ inefficiencies and trends• Arbitrated between 10+ marketplaces with very heterogeneous technicalitiesUse of : Pytorch, Tensorflow, Keras, transformers, Scrapy, Beautiful-Soup, Request, Flask

Worked on a software that analyzes data from any written medical source, to sum up patients’ cases and suggest medical diagnoses and potential alerts for health institutes• Used complex models including RNNs of GRUs/bi-LSTMs, own-trained word embeddings, transfer learning, multi-task learning, and zero-shot learning• Revamped and optimized our prototype’s data pipeline from the processing of the raw scans and models’ training to the actual predictions made inside our app• Read numerous research papers pertaining to diverse aspects of our models and domain of application to exploit the latest advances and most relevant techniques• Handled confidential data in a highly regulated field with no margin for error• Highly sparse, noisy and biased data and labels: 68K+ labels (ICD-10) with sometimes no incidence occurring for the last 10 yearsUse of : Tensorflow, Keras, Gensim, NLTK, Docker, GPGPU computing, Linux, SQL, Requests, Flasks

Project:• Built a generalist AI for Atari games combining reinforcement learning and deep learning using a technique pioneered by Goggle DeepMind in December 2013 called Deep Q-Learning => use of python's library Theano and RL-Glue to create the agent and interact with the Arcade Learning Environment => use of Lasagne on top of Theano and cuDNN for accelerated resultsA web site containing information about the procedures and results is available at https://deepqlearning.wordpress.com/• Solved a Kaggle competition with the use of Deep Learning => use of Keras and Lasagna libraries on top of Theano• Conducted a series of interviews with experts evolving in the Data Science industry and creation of a blog containing the published interviewshttp://datainterviews.com/Report:• Analysis of the Data-Science-as-a-Service ecosystem• Analysis of the evolution of the Deep Learning architectures• Analysis of the past evolution and prediction for the GPU market

Projects:• Developed sentiment analysis classifiers combining NLP methods with information extraction and machine learning algorithms in Python and R. => acquisition of insights regarding the emotional status, basic psychology and human values of the audience of one's Twitter accountsee "Explore your Twitter Audience Psychology" on http://labs.wisemetrics.com/• Developed a web tool to analyze users’ Twitter audience in order to rank them amongst other users and give insights about their followers. => development and smoothing of a ranking strategy and optimization of sampling and ranking methods to get an quick online grading of one's Twitter account• Developed internal tool to hunt for interesting hash-tags on Twitter and forecast the evolution of their future usage and impact on Twitter. => development of an hash-tags scoring strategy, sampling and forecasting of the score time-series• Designed and implemented innovative add-ins to the existing products.

• Studied the penetration and market share of very high speed broadband regionally• Implemented statistical tools to report on evolution of key network performance indicators • Analyzed big databases (10M+ lines)