Marc Benzahra

DinMo is a platform designed for growth teams so they can leverage their customer data in all their activations platforms without relying on engineers or analysts. Our mission is to simplify data access, eliminate complexity, and promote a data-driven culture that empowers teams to make informed decisions and fuel growth. I have the opportunity to build the Machine Learning Platform from scratch using Snowflake Container Services. After phases of contribution to product framing and model prototyping, I deployed models as products in DinMo's system such as LTV/Churn scoring and Customer/Product Recommendations. This work is done in collaboration with a team composed of a data scientist, a data engineer, a product manager and a backend engineer.

WedR is a consulting company, part of Stanwell. We build projects for our clients and help them lay out foundations of Software, Data and AI Initiatives. I mainly did software development, data science, machine learning, data engineering, MLOps/DevOps. Client was Société Générale Factoring, as a Software Engineer for WedR, mission was to refactor and put a Machine Learning model into production using RESG private cloud and file transfer tools. This model's goal was to predict factoring customer's risk of being in debt. I also animated internal presentations at WedR and led a LLM project based on Snowflake and BLOOM.

Applied Data Science and Natural Language Processing techniques to readability assessment. Supervised learning to automatically compute text readability in English and French. Research work on word and subword language modeling as an unsupervised task to assess text readability with surprisal metrics.

Applied Data Science and Natural Language Processing techniques to readability assessment. Supervised learning to automatically compute text readability in English and French. Research work on word and subword language modeling as an unsupervised task to assess text readability with surprisal metrics. Worked in a team of 3 engineers for production tasks and solo during model building.

Research work on word and subword language modeling as an unsupervised task to assess text readability with surprisal metrics. Built a universal assessment of texts readability, universal in two senses: - multilingual, need for unlabeled data as annotated data is scarce or unstructured in text readability assessment - diachronic proof, language semantic is affected by time, word senses evolve through history, an initial goal of my subject was to handle these effects As literature suggested, reading times are correlated to words surprisal and reading times are correlated to cognitive load during silent reading. Thus, we hypothesised that text surprisal would be a good indicator of text readability. I worked on surprisal metrics computed with language models (LSTM, AWD-LSTM, N-Grams, Transformers, GPT-2 and 3, BERT) to produce readability assessments of texts.

We help modern sales & marketing teams to reach their business goals with a unique data-driven sales engagement software. Engage contacts with automated email sequences, identify business opportunities and grow a predictable revenue pipeline with our sales engagement software. Find new cold leads, turn them into customers. I built several natural language processing projects during my stay at Datananas such as: - A company similarity detector using a feed forward neural network and a large collection of company entries, the first version was built using non labelled data, with genetic algorithm principles (mutation and crossover with heuristics to evaluate company entries promixity) and evolving neural networks (NEAT). However, we built a labelled dataset and a simple feed forward neural net was enough to solve the task. - A Job titles clustering using job titles and descriptions with an unsupervised algorithm (k-means) - An improved postal location detector that I did during my internship in the company (recurrent neural network, word2vec embeddings, supervised learning). - An email redaction scoring algorithm using linear and polynomial regression to measure how well a prospection email is written according to expert features (supervised learning). - Helped improve language detection in emails and the existing email classification system using a naive bayes algorithm for classification. - Developed a postal location named entity recognition algorithm which was rule-based. I also proposed prototypes to replace our n-gram units by continuous bag of words with word2vec. I worked in a tech team of 8 members, we were 2 working on natural language subjects.

Smartly.ai is a web platform to develop chatbots, visualise and arrange the dialog flow of your chatbot with existing pretrained intents which are also customizable. Logic can be programmed for each dialog step with custom code and integrate any API. The bot flows developed on the platform are deployable as text and speech chatbots on platforms such as Messenger, SMS, Alexa, Google Home and more. I designed, developed and maintained chatbots on our platform smartly.ai for customers such as RATP, SNCF, Neopost, Orange, SEB. Took part to the back-end development of the web platform. Benchmarked most of the chatbot tools available in the industry landscape of 2016: - api.ai (former company of Dialogflow from Google) - luis.ai (intent detection from Microsoft) - Watson, a large toolkit from IBM, including some natural language understanding tasks that we had to tackle - Rasa NLU, open source tools for natural language understanding Lastly, I was in charge of the second version of the dialog engine in Node.js from the first revision done in C++ few years ago (this work involved generation and walk of dialog graphs). We were 6 members in our tech team, I was both with the chatbot development team and the back-end development team.

Knowledge Inside solutions have enabled customers to model innovative systems (electric & autonomous vehicles) and produce their specifications, to respond to calls for tenders twice as fast, to model services involving many players, to meet standard operational safety requirements (railway signaling), to reverse engineer existing toolchains and more. During this internship I implemented the MIC data model (https://tel.archives-ouvertes.fr/tel-01168567/document) from an existing work in our main software: arKItect. I made a library so that our developers and users were able to use the data model. Moreover, I also provided few interfaces that used this data model to implement features in our software such as the import and export of projects from our software arKItect to system simulation software like AMEsim and Simulink. I worked in a tech team of 6 members split in Russia and France, I was in charge of the MIC project lead by our CTO and with the mentoring of both the french and russian teams.

Reworked a business intelligence web app: ForecastTest. Edition and visualisation of tests forecasts. The new version we developed delivered a performance boost for both the front-end (UI improvements) and back-end (execution speed). We carried the project as a team of 2 members including me, in the internal tools and support division of Alcatel-Lucent.