Wiremind - Machine learning researcher (internship)

Our algorithms are divided in 2 parts:

• A modelling of the unconstrained demand using ML models (e.g. deep learning, boosted trees) trained on historical data in the form of time-series

• Constrained optimizations problems solved using linear programming techniques

You will be joining a team shaped to have all profiles necessary to constitute an autonomous departement (devops, software and data engineering, data science, AIML, operational research). 

There, under supervision of a wiremind tutor and researchers from the LPSM (https://www.lpsm.paris/) you will research a new avenue of modelisation. 

In this project, we propose to introduce new deep learning architectures inspired for instance from [2] to propose parametric time series generative modelling for the unconstrained demand with price elasticity. These approaches are based on general state space models where the observations are random functions of the latent states. In such a setting, inference procedures require the posterior distribution of the latent states given the observations that are intractable. Traditional Markov Chain Monte Carlo and Sequential Monte Carlo methods are too computationally intensive to be used in the context of this project so that we propose to explore variational inference and VAE methods. Such approaches would allow to introduce approximate posterior distributions and predictive distributions for the unconstrained demand with price elasticity. These distributions can then be used to provide uncertainty quantification metrics. The main objective of this project is to propose new models that specifically takes into account decreasing monotone constraint between demand and price. This is an open problem both from practical and theoretical perspectives. One solution is to introduce specific conditional likelihood to allow dynamic pricing with monotonicity constraint. We will also investigate the introduction of constraints in the reconstruction loss of the variational model to satisfy the required monotonicity.

Technical stack:

• Python 3.11+

• Argo over an auto-scaled kubernetes cluster for orchestration

• Druid as datastore

• Common ML libraries: tensorflow, lgbm, xgboost, pandas, dask, dash, mlflow

• Gitlab for continuous delivery

ABOUT US

Since 2014, Wiremind has been providing traffic and revenue optimisation solutions for the transport and sports-event industries. Thanks to our SaaS applications, we combine UX, Software and Data Science and collaborate with leading international clients such as SNCF, PSG or even WestJet and Qatar Airways.

We are working on a range of projects, including forecasting and pricing rail passenger demand, optimising aircraft loading in 3D and managing ticketing for sporting events in real time. Our applications are used on a daily basis by hundreds of users, particularly rail and airline companies, in many countries and on several continents.

With over 70% of our team made up of tech professionals, we are growing by 100% every 18 months. Our business model is based on software-as-a-service solutions with long-term contracts, enabling us to maintain high profitability without the need to raise capital.