• 🔭 I’m the founder of https://www.seocopilot.fr : Helping compagnies to improve their SEO organicaly
• 🌱 I’m currently learning about Data Science / AI / Deep Learning / Machine learning (Never ending process)
• 👯 I’m looking to collaborate on everything that has a meaningful purpose when I have time !
• 📫 How to reach me: https://www.linkedin.com/in/smougel/
• 😄 Pronouns: He/His

I've a background of 20 years as Software Engineer (Fullstack dev) and now I add a new card to my set as Data Scientist / AI Engineer. I love to learn new things about Deep learning / CNN / Sequence Models (Many thanks Andrew NG & Coursera)

• 💻 Programming languages : Python, Javascript, Php
• 🗂️ Databases : Maria DB / MySQL / Redis
• 💍 Front : React / Redux / Css / HTML 5
• 🔧 Web frameworks : Symfony / Laravel / Code Igniter
• 🔩 DB frameworks : Active Records, Doctrine ORM
• ⚙️ Backend : Writing of Workers & Daemons
• ⌚ Load & Queue management : Beanstalkd
• 📊 DataViz : Matplotlib , plotly, seaborn
• 🧪 Data science : Pandas, numpy, scikit learn, Tensorflow, Pytorch, keras

I will open source them as soon as possible.

Goal : Helping Paris city to become a smart city.

Optimization of tree maintenance

Data source : opendata.paris.fr

My work : https://github.com/smougel/eda_paris_tree (Notebook & Presentation)

Goal : Analyse and find healthy products / Inform people about nutritional metrics

My work : https://github.com/smougel/eda_open_food_facts (Notebook & Presentation)

Build a model to detect people able to repay their loan... or not...

💡 Process :

• exploratory data analysis
• data cleaning
• feature engineering
• sampling / train & test split
• model training : SVM, Neural Networks, Logistic Regression, Random forest
• variable importance evaluation with lime

📏 Metrics : Precision / Recall / F1-Score

🪛 Hyperparameter tuning with grid-search and cross-validation

Tech used : Python, Scikit-learn, Matplotlib, Seaborn Project : https://github.com/smougel/credit_scoring/tree/master (Notebook)

Unsupervised learning task

Dataset : Olist.com

Gain insights about user behavior and discover buyer characteristics

💡 Process :

• exploratory data analysis
• data wrangling
• feature engineering
• dimensionality reduction : Principal Component Analysis
• clustering : k-means, db-scan
• elbow method
• High-dimensionality vizualisation : T-SNE, U-Map
• Analysis of cluster stability

📏 Metrics : ARI Score

😡 Customer dissatisfaction discovery 🪄 Automatic photo classification : Menu card, Food picture 🍝, Outdoor picture

Dataset : https://www.yelp.com/dataset

💡 Process :

• exploratory data analysis
• data wrangling for NLP (🤬 stop words , lemmatization, stemming, tokenization)
• data wrangling for photos (contrast normalization, resizing, noise filtering)
• Topic discovery : LDA
• Convolutional neural networks
• Regularization (Dropout)
• Exploration of filters learned by the CNN (Thanks to francois Chollet)
• Use of Yelp API

Tech used : Tensorflow, NLP, Sequence model, LSTM, CNN, Keras, Open CV

Dataset : https://www.kaggle.com/kazanova/sentiment140 (1.6M Tweets) Goal : Sentiment analysis from tweets. Benchmark with Microsoft Azure Sentiment Analysis.

💡 Process :

• exploratory data analysis
• data wrangling (lemming / stemming / tokenization)
• modelization (Basic to advanced : logistic regression, TF-IDF, LSTM)
• benchmark with Azure Machine Learning Services

🪛 Hyper parameter tuning 📏 Metrics : F-Beta Score

Tech used : Word embeddings (Word2Vec & FastText), Tensorflow & Keras

Dataset : Cityscape

💡 Process :

• exploratory data analysis
• data wrangling (Picture to binary mask)
• data augmentation (Random cropping, flipping, mirroring)
• modelization (Basic to advanced : Fully connected layers to U-Net architecture)
• ☁️ training in the cloud (w/ Microsoft Azure : compute instance provisionning)
• Model serving via Flask API hosted on Microsoft Azure

🪛 Hyper parameter tuning 📏 Metrics : Jaccard index

Tech used : Tensorflow, Keras, CNN, U-Net, Flask, Azure Services

Dataset : News Portal User Interactions by Globo.com https://www.kaggle.com/gspmoreira/news-portal-user-interactions-by-globocom#clicks_sample.csv

💡 Process :

• exploratory data analysis (w/ t-SNE visualization of news embeddings)
• data wrangling
• modelization : content filtering and collaborative filtering
• ☁️ training in the cloud (w/ Microsoft Azure : compute instance provisionning)
• Use of serverless Azure Function for model serving / Azure Storage
• Integration with a node js mobile app

📏 Metrics : Similarity measure (dot product, cosine)

Tech used : Tensorflow, Sparse Tensor, Matrix factorization, Azure Services

Dataset : Microsoft frames dataset (Dialogs between two humans via a chat interface) https://www.microsoft.com/en-us/research/project/frames-dataset/

💡 Process :

• exploratory data analysis (w/ t-SNE visualization of news embeddings)
• data wrangling
• LUIS Training
• Integration w/ Microsoft bot framework

📏 Metrics : Similarity measure (dot product, cosine)

Tech used : Microsoft LUIS, Microsoft Bot Framework, Azure application insight, Unit Testing