Experience

QUANT AI Lab

• France
• Information Technology & Services
• 1 - 100 Employee

• AI/NLP Engineer

  • Feb 2022 - Present

  Building NLP Products around LLMs using Transformers, PyTorch, Langchain, Pydantic, FastAPI & MongoDB. Building NLP Products around LLMs using Transformers, PyTorch, Langchain, Pydantic, FastAPI & MongoDB.



Pépite France - Le réseau des Étudiants-Entrepreneurs

• France
• Higher Education
• 100 - 200 Employee

• Étudiant Entrepreneur

  • Sep 2021 - Jun 2022

  I joined this program to live an entrepreneurship adventure and learn through the way skills revolving around marketing, communication, public speaking and the construction of business plans. I gained the National Student-Entrepreneur Status (SNEE), which allows students to build and develop their entrepreneurial projects and receive support. I joined this program to live an entrepreneurship adventure and learn through the way skills revolving around marketing, communication, public speaking and the construction of business plans. I gained the National Student-Entrepreneur Status (SNEE), which allows students to build and develop their entrepreneurial projects and receive support.



LightOn

• France
• Software Development
• 1 - 100 Employee

• Research Intern in NLP

  • Mar 2021 - Aug 2021

  Published paper: https://arxiv.org/abs/2109.11928 - Exploration the use of random parameter models in Language models : Transformer (GPT-2), Reservoir and Doped), compared them and chose Doped because of its performance and the simplicity of its architecture. - Experimentation on the order, number and structure of layers of each model (Transformer and Doped Transformer models with 355m params and 774m params) and the simplicity of its architecture. - Evaluation of Transformers… Show more Published paper: https://arxiv.org/abs/2109.11928 - Exploration the use of random parameter models in Language models : Transformer (GPT-2), Reservoir and Doped), compared them and chose Doped because of its performance and the simplicity of its architecture. - Experimentation on the order, number and structure of layers of each model (Transformer and Doped Transformer models with 355m params and 774m params) and the simplicity of its architecture. - Evaluation of Transformers (small and medium GPT2 / doped GPT-2) on downstream tasks of NLP (Orangesum and FLUE) Show less Published paper: https://arxiv.org/abs/2109.11928 - Exploration the use of random parameter models in Language models : Transformer (GPT-2), Reservoir and Doped), compared them and chose Doped because of its performance and the simplicity of its architecture. - Experimentation on the order, number and structure of layers of each model (Transformer and Doped Transformer models with 355m params and 774m params) and the simplicity of its architecture. - Evaluation of Transformers… Show more Published paper: https://arxiv.org/abs/2109.11928 - Exploration the use of random parameter models in Language models : Transformer (GPT-2), Reservoir and Doped), compared them and chose Doped because of its performance and the simplicity of its architecture. - Experimentation on the order, number and structure of layers of each model (Transformer and Doped Transformer models with 355m params and 774m params) and the simplicity of its architecture. - Evaluation of Transformers (small and medium GPT2 / doped GPT-2) on downstream tasks of NLP (Orangesum and FLUE) Show less



Sorbonne Université

• France
• Higher Education
• 700 & Above Employee

• Student Job at Fablab / PartiCitaE

  • Sep 2019 - Aug 2021

  PartiCitaE : - Design, writing, animation and reporting - Video editing and production - Moderator on social networks Fablab : - Computer support and help in the use of 3D printers, laser cutter, heat press machine, PartiCitaE : - Design, writing, animation and reporting - Video editing and production - Moderator on social networks Fablab : - Computer support and help in the use of 3D printers, laser cutter, heat press machine,



LIP6 - Laboratoire d'Informatique Sorbonne Université/CNRS

• France
• Research Services
• 1 - 100 Employee

• Research intern in NLP

  • Jun 2020 - Aug 2020

  - Training language models (Bert/Transformer) and evaluating them. - The model directly used 700gb of processed data during the training of the model. - However, the server (4 x Titan X and 392GB Ram) could not support efficiently all the data size . The solution found was the following: the unprocessed data weighed 80GB, we used the unprocessed data directly in the model, and at each layer of the model we built what we needed and deleted what we no longer needed. - Training language models (Bert/Transformer) and evaluating them. - The model directly used 700gb of processed data during the training of the model. - However, the server (4 x Titan X and 392GB Ram) could not support efficiently all the data size . The solution found was the following: the unprocessed data weighed 80GB, we used the unprocessed data directly in the model, and at each layer of the model we built what we needed and deleted what we no longer needed.



Laboratoire Images, Signaux et Systèmes Intelligents - UPEC

• Région de Paris, France

• Research Intern in Computer Vision & Deep Learning

  • Jun 2019 - Sep 2019

  The Subject was "The Recognition of emotions from images and videos of people's faces" - Trained A CNN for facial_expression_recognition - Trained GANs and DCGANs Models for data augmentation - Published a paper: https://www.researchgate.net/publication/348545766_Multi-facial_patches_aggregation_network_for_facial_expression_recognition_and_facial_regions_contributions_to_emotion_display The Subject was "The Recognition of emotions from images and videos of people's faces" - Trained A CNN for facial_expression_recognition - Trained GANs and DCGANs Models for data augmentation - Published a paper: https://www.researchgate.net/publication/348545766_Multi-facial_patches_aggregation_network_for_facial_expression_recognition_and_facial_regions_contributions_to_emotion_display



Hack iT

• Algeria
• Software Development

• 1er Prix Hackathon Smart City Algiers

  • Apr 2018 - Apr 2018

  Winner of Hack it - smart cities Algiers Winner of Hack it - smart cities Algiers



El Djazair Information Technology (ELIT)

• Algeria
• IT Services and IT Consulting
• 200 - 300 Employee

• Back End Dev Intern

  • Jul 2017 - Aug 2017

  Participated in the development of a pharmaceutical production management application using JEE (JSF) and PostgreSQL. Participated in the development of a pharmaceutical production management application using JEE (JSF) and PostgreSQL.