About Me

Hello! I'm Audrey DJOUPE PENE, a passionate Data Scientist and Machine Learning Engineer. I specialize in developing innovative solutions using cutting-edge AI technologies.

Skills

Machine Learning & Deep Learning
Data Analysis & Visualization
Python, TensorFlow, PyTorch
Natural Language Processing
Statistical Analysis

WORK EXPERIENCE

BpiFrance

Role: Data Scientist Developer

Development of a Chatbot to Improve Financial Strategies

Data collection and structuring from various sources to prepare datasets.
Selection and fine-tuning of Llama models to meet the client's specific needs.
Backend Development: Integration of AI models into a secure and scalable architecture.
Optimization of CPU and memory resources to ensure optimal performance.
Security of sensitive data and secure integration into a Docker environment.
Deployment of the system in production on an ultra-secure VPS.

GROUP VERDON

Role: AI Engineer

Development of a personalized chatbot for the company

Analysis, processing and cleaning of the company's unstructured database for better use of information.
Design and implementation of structured and optimized SQL databases for storing and managing data in vectors.
Development of a custom intelligent chatbot, trained on internal company data using advanced LLM models (Llama and DeepSeek) locally.
Chatbot optimization to ensure relevant responses using RAG and Search Engine methods.

TOTALEnergies

Role: Data Scientist Consultant

Detection of emotion, gender and age on an embedded system to assess TotalEnergies customer satisfaction

Carrying out an in-depth benchmark of existing data and solutions to align the POC with TotalEnergies' specific needs
Selection and analysis of relevant images for training 3 detection models (age, gender and emotion)
Training and optimizing each detection model on data for real-time operation
Integration of the 3 detection models into the embedded system for extracting the age, emotion and gender of customers, then producing dynamic statistics on the results.

CAPGEMINI ENGINEERING

Role: Data Scientist Consultant

Development of a model to assist in the diagnosis of patient illnesses

Carrying out a benchmark of existing data and solutions to align the methods to be used with the needs of the project.
Collection and implementation of an ETL pipeline for processing more than 10,000 radiographic images improving their quality using Azure Data factory, Azure Synapse Analytics, Azure Data Lake Gen 2, Databricks, Azure Key Vault.
Data visualization and rebalancing for analysis to optimize model training.
Development of a high-performance AI model (87% accuracy) integrated into a web application for doctors, with a CI/CD pipeline for fast and reliable deployments (Docker, Git, AWS), ensuring fast and continuous updates.

From Loops to Lightning - How Transformers Outran RNNs

Mar 21, 2025

This article gives an in-depth overview of the Transformer architecture, which has revolutionized natural language processing. It focuses on attention blocks, the key component of the model that establishes parallel and contextual connections between words in a sentence.

Recurrent Neural Networks uncovered — The power of memory in deep learning

Mar 12, 2025

This article talks about how deep learning has transformed various fields, highlighting the strengths and limitations of Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). It explains that RNNs are designed to process sequential data by maintaining memory of previous inputs, making them ideal for tasks like natural language processing and speech recognition. The article also discusses advanced RNN variants like LSTM and GRU, which improve learning of long-term dependencies. Finally, it mentions the evolution toward Transformer models, which have become the new standard for handling complex sequence data efficiently.

Deep Learning basics for video — Convolutional Neural Networks (CNNs) — Part 2

Mar 1, 2025

This article explains different activation functions used in neural networks, such as Sigmoid, Tanh, and ReLU, highlighting their advantages and limitations. It describes the vanishing gradient problem, which slows down learning in deep networks due to very small gradients. The article also covers how backpropagation adjusts weights using gradients to improve model predictions. Finally, it explains pooling layers and fully connected layers, essential components in convolutional neural networks for feature reduction and decision making.

Recent Projects

Text summarizer MLOps

Transformers

Tensorflow

Torch

Matplotlib

Fastapi

Generates a text summary using a Huggingface model trained on a set of texts. Deploy the summary generation process on a platform using AWS tools

Dance style classification

Pytorch

Opencv

Alphapose

Tensorflow

Pandas

Creation of a dance style classifier (afrobeat, hip-hop, classic.) from collected and processed videos, using convolutional neural networks on GPUs. Automatic extraction of key kinematic features and pose estimates via Mediapipe and AlphaPose. Deployment on a Web application for real-time classification.

Neural style transfert

python

tensorflow

tensorflow-hub

matplotlib

numpy

Application of neural style transfer to generate images combining the content of a photo and the style of famous works of art. Using PyTorch and VGG pre-trained models to isolate and recombine content and style features.

3D Generative Adversial Model with VAE encoders

Python

Torch

Numpy

Matplotlib

PIL

The project aims to implement and compare GAN, VAE-GAN and MV-VAE-GAN models for 3D shape generation. These models are trained to generate new shapes from latent spaces.

Chatbot fine tuning from scratch

Tensorflow

Torch

Flask

Transformers

NLp

Colorama

Development of a conversational chatbot based on an open-source Large Language Model (LLM) to automatically generate articles on various topics. Analysis of the current capabilities and limitations of LLM models for the production of autonomously written content.

Federated Learning

Tensorflow federated

Numpy

Pandas

This project aims to enable the continuous improvement of clinical prediction models used in hospitals, while fully respecting the confidentiality of patient data thanks to federated learning.