
AI is analyzing your overall score…
Identifying your key strengths…
Evaluating your skill match against the job requirements…
Assessing your cultural and operational fit
Co-Founder & CTO @Klineo | Angel Investor in Tech, Health & Impact | Hiring Tech Talent
As a generalist engineer with expertise in Machine Learning and Computer Science, I build tech solutions that tackle real-world challenges and improve lives. At Klineo, we use AI to connect cancer patients with clinical trials tailored to their needs—making cutting-edge treatments more accessible. I also invest in and advise early-stage ventures in tech, health, and impact, backing founders who turn innovation into meaningful change. My mission: leverage technical expertise to support ideas that make the world better, one solution at a time.
Harvard University
Master of Science (M.Sc.), Computational Science & Engineering
January 1, 2015 – January 1, 2016
École Polytechnique
Engineer's degree, Entrepreneurship and Big Data
January 1, 2012 – January 1, 2015
Lycée Sainte Geneviève
Mathematics, Physics, Engineering Science and Chemistry
January 1, 2010 – January 1, 2012
Klineo
CTO & Co-Founder
July 1, 2022 – Present
Iktos
Principal Scientist, Group Leader, AI
December 1, 2020 – July 1, 2022
Stealth
Angel Investor
November 1, 2019 – Present
Le Wagon
Track Owner for Deep Learning
November 1, 2019 – December 1, 2022
Greater Paris Metropolitan Region
Iktos
Senior Data Scientist
March 1, 2019 – December 1, 2020
AnotherBrain
Senior AI Engineer
January 1, 2018 – March 1, 2019
Paris Area, France
Salesforce
Senior Data Scientist
May 1, 2017 – October 1, 2017
Salesforce
Data Scientist
July 1, 2016 – April 1, 2017
tinyclues
Data Scientist Intern
March 1, 2015 – August 1, 2015
Paris
Floating Apps, Inc.
Summer Intern
July 1, 2014 – August 1, 2014
New York City
High Mountain Military School
Cadet Officer
December 1, 2012 – April 1, 2013
Chamonix
Log-Structure Merge Tree for Key-Value Stores
March 1, 2016 – May 1, 2016
This is an individual project about designing and implementing the basic structure of an LSM tree for reads and writes in a parallel way. Indexing with LSM Trees is a core component of many modern key-value stores as this storage-aware data structure is designed for low-cost indexing for a file experiencing a high rates of record inserts and deletes over an extended period. The deliverable is a persistent LSM Tree with an update rate about 0.25 M per second and a read rate about 22.5K per second on flash storage. More information on the online paper: Development in C. Code available on github.
Time Series Database and Library
February 1, 2016 – May 1, 2016
This is a group project for the development and the documentation of a persistent time series database with its analysis library. The database enables to store with persistence time series with metadata through the REST api provided. The library enables to run analysis on time series and to augment the metadata (mean, standard deviation, euclidean distance computation...). This library provides a sophisticated similarity measure between 2 time series based on an iSAX tree (http://www.cs.ucr.edu/~eamonn/iSAX_2.0.pdf). I was particularly involved in the persistence design and in the analysis library. Development in Python 3. Code available on github
Human Activity Recognition Using Time Series Analysis
January 1, 2016 – May 1, 2016
This is a group project about human activity recognition using various Time Series methods. We applied on sensor data the following methods: Conditional Random Fields, Hidden Markov Models and Max-Entropy Markov Models. We reached an accuracy of 97% using CRF. Development in Python. Code available on github.
Harvard Student Computational Challenge: Human Flow Analytics
January 1, 2016 – Present
With the Internet of Things becoming a reality, we increasingly find sensor devices being deployed to monitor the movement of people in commercial and institutional buildings. Information from these sensors is useful in automating various operations in the buildings ranging from control of cooling systems, to security monitoring, to emergency evacuation planning. However, in practice, sensor devices can be highly unreliable and are prone to frequent failures. As a result, the data collected from them are often incomplete and noisy. I took part to this challenge with a classmate. We were provided highly incomplete data from sensor readings of human movement in a building. The challenge was to accurately predict the missing data by exploiting the recurring temporal and spatial patterns in the way people move around within the building. We split the work into two parts. First, we constructed a weighted neighborhood graph to build a first model with a nearest neighbors approach. Second, for each sensor, we built a set of features, using amongst other things measures from its neighbors, averaged with regards to different time frames. We then applied and combined penalized linear regressions. The most exciting part was the live leaderboard we led during the entire competition and where we could see the improvements of our models. Check out our code on Github
Memory Network For Question Answering
January 1, 2016 – May 1, 2016
This is a paired project about about non-factoid question answering. We first implement a count-based baseline where we realized that solving this kind of problems requires to process input data and analyse them similarly to the memory process. We thenchose to implement a method based on memory networks and obtained competitive results with state-of-the-art methods. Development in Torch, using the libraries rnn and graphnn. Code available on github.
Fine-Grained Recognition using Parts-based One-vs-One Features
September 1, 2015 – December 1, 2015
This is an individual project for the CS 283 Computer Vision course at Harvard University. I decided to tackle a category-level recognition problem, where the objective is to recognize any instance of a particular general class. I implemented the Part One-vs-One Features (PooF) method for fine-grained recognition, developed in http://thomasberg.org/papers/poof-cvpr13.pdf. This method automatically learns a large and diverse set of features specialized in the discrimination between two particular classes based on the appearance at a particular part. I applied it to a birds dataset containing 200 unique species in about 11 800 images (the CUB-200-2011) and on a whales data set from a Kaggle competition. The development was done in Matlab using the Computer Vision Matlab Toolbox and the VLFeat library.
Category Classification of Yelp Restaurants
September 1, 2015 – December 1, 2015
This is a paired project for the CS 281 Advanced Machine Learning course at Harvard University. The objective is to improve the categorization system of Yelp businesses using machine- learning techniques. To achieve our goal, we first extracted significant information out of the text reviews posted on Yelp using dimensionality reduction methods. We coded in Python our own Latent Dirichlet Allocation (LDA) using online variational inference , and benchmarked it with the LDA using a Gibbs sampling method and with the Non Negative Matrix Factorisation. We then applied both supervised and unsupervised learning methods using the sickit-learn library. First, we used one-vs-all classifiers on the text reviews combined with characteristic features of businesses to predict the category labels. Second, we clustered the entries using a community detection algorithm on the network of restaurants. This roadmap was applied on the restaurants in Las Vegas and provided promising results which could be used as the first step to an automated categorization system on the Yelp website. Feel free to check our report online!
Parallel Latent Dirichlet Allocation
September 1, 2015 – December 1, 2015
This is a group project for the CS 205 Computing Foundations for Computational Science course at Harvard University. The objective is to implement an LDA package in Python that gathers the two most common inference methods: the gibbs sampler and the online variational inference. We took advantage of the Cython and Multithreading modules to build an optimized and parallel algorithm. Our deliverable is an open-source library: fully documented, easy to install and much faster than the current tools. To tackle this problem, we had to deal with optimizing performance in Cython and parallel design. Once that was completed, we needed to bring the different methods together into a functional package.
Perceptron Ventures: Predicting start ups’ success
September 1, 2015 – September 1, 2015
This is a group project for the CS 109 Data Science course at Harvard University. The objective is to predict the startups’ success based on the large amounts of data available online. We gathered data for nearly 7000 companies from the Crunchbase API that were founded prior to 2011, with information on location, sector, the company's founders, funding rounds, acquisitions IPO’s… Our goal was to come up with a definition for what constitutes a successful company based off some of these features, and to accurately classify the companies. As we analyzed these companies to determine whether we should fund a startup, we also looked at the problem through the lens of Profit and ROC curves, as funding a company that is ultimately unsuccessful has a much different cost than missing a company that winds up successful. Finally, we wanted to approach the data again from an entirely different direction - seeing if we could cluster certain startups together through a similarity measure. We used this to see which companies were connected, and tried to glean useful information from these connections. This project was done with Python, taking advantage of the libraries Pandas and scikit -learn.
Ray Tracer
March 1, 2014 – June 1, 2014
In a computer graphics project, I designed a ray tracer algorithm capable of building an image from a scene composed of spheres interacting with light. Ray-tracing is a rendering technique for generating an image by tracing the path of light through pixels in an image plane and simulating the effects of its encounters with virtual objects. The projects called into play both my analytical skills to compute the effects of reflection and shadows on the spheres and my Java proficiency to come up with a smart program and a user-friendly interface.
Game Theory
Coursera
June 24, 2026 – Present
Big Data Analysis with Scala and Spark
Coursera
June 24, 2026 – Present
Functional Programming Principles in Scala
Coursera
June 24, 2026 – Present
Machine Learning
Coursera
June 24, 2026 – Present
Parallel programming
Coursera
June 24, 2026 – Present
Computational Neuroscience
Coursera
June 24, 2026 – Present
Cultural Fit Analysis
The candidate's diverse project portfolio, ranging from academic research to startup ventures and angel investing, indicates a strong drive for innovation and a broad interest in technology and its applications. Their experience in health tech (Klineo, Iktos) and impact investing suggests alignment with mission-driven organizations. The role as a Deep Learning Track Owner at Le Wagon also shows a commitment to knowledge sharing and community building. This profile suggests a good cultural fit for dynamic, research-oriented, or impact-focused environments.
Soft Skills & Operational Fit
The candidate's experience as a CTO, Group Leader, and Track Owner for Deep Learning suggests strong leadership, mentorship, and communication skills. Their involvement in multiple group projects and a startup indicates an ability to collaborate and drive initiatives. The military background also implies discipline and structured thinking. These traits align well with senior operational roles requiring both technical depth and team management.