Romain Zimmer

First ML Engineer hired to design and build a discount optimisation system from scratch to help retailers reduce waste and increase profits. Including: - Setting up and maintaining data pipelines (DBT, Snowflake) - Developing ML models to predict discount elasticity - Developing an optimisation system to decide which discount to apply and when - Developing and maintaining ML pipelines (Metaflow) - Working with other teams to improve data quality and availability - Developing a small experimentation framework and running AB tests in stores - Working closely with our PM to improve the product I'm also actively contributing to the improvement of ML practices at Too Good To Go.

Yokai.ai: - Python code refactoring - Migration from Flask app to Celery app with Redis backend - Implementation of integration tests (Docker Compose, MinIO, Pytest, Celery, Redis) running in Github worflow - Migration from Kubernetes on Google Cloud Platform to Serverless + S3 + Cloudfront + Terraform on AWS - Implementation of a simple authentication feature using Lambda Authorizer, DynamoDB and Basic Auth. - Development of a slack bot to validate images generated by a ML pipeline - Deployment of a Triton Inference Server on AWS Sagemaker Multi-Model Endpoint with Terraform and a Gradio interface with the Serverless Framework

Joined Inria's Startup Studio and received 1 year of funding to build Sievable, a collaborative filter-based search engine powered by machine learning.

Building a collaborative filter-based search engine powered by machine learning, including: - training of a neural network to predict whether a natural language filter matches an item description or not - design of a scalable search engine system with search result candidates selection using clustering - designed a custom MLOps pipeline to automatically pull our datasets from different sources, train our model, update our items embeddings and deploy everything - developed our own search engine platform allowing users to search, index new content and contribute by annotating data - created our own cryptocurrency on the Solana blockchain to reward contributors Using: PyTorch, FAISS, MySQL, Redis, GraphQL, React, Solana Javascript API, RabbitMQ, Terraform, Docker, Kubernetes, Cloudflare Deployed on Scaleway.

R&D on generative models for speech synthesis Development of a Python library to estimate ML pipelines performances on ARM processors

Development of Spiking Neural Networks (SNNs) for audio analysis. Spending my time between the Center for Brain and Cognition Research (CerCo) and the Institute of Computer Science (IRIT) of Toulouse, two labs of the French National Centre for Scientific Research (CNRS). Modelling the dynamics of SNNs by a Recurrent Neural Network (RNN) trained with surrogate gradient methods. The final model achieved comparable performances as traditional deep learning models while using a binary activation function with low firing rate.