Leading a team of scientists and engineers in Amazon's central Economics and Science organization with focus on probabilistic & causal inference for demand modeling and elasticity estimation.

Science lead of a team developing an internal probabilistic programming language with a strong focus on ease of use and robustness. Customers are business teams that aim to combine domain expertise, business constraints, and data.

Managing teams to build ML services including a multi-armed bandit service for online A/B testing, active-learning based labeling (now AWS SageMaker Ground Truth), and hyper-parameter optimization (now part of AWS SageMaker HPO).Selected Publications:- C. Vernade, A. Carpentier, G. Zappella, B. Ermis, M. Brückner. Contextual Bandits under Delayed Feedback. Technial Report, 2017.

Building backend services and libraries to enable the usage of Machine Learning across Amazon and AWS.Achievements:- Design and implementation of a Java ML library used in AWS Machine Learning and other services. The library implements a variant of stochastic gradient descent for various learning tasks; it scales to millions of data points and data attributes.- Design and implementation of a Scala/Spark-based library for incremental variational inference to learn probabilistic models, e.g., for Thompson sampling.Selected Publications:- M. Brückner, D. Blick, Optimized training of linear machine learning models, US Patent 10,318,882- L. P. Dirac, M. Brückner, R. Herbrich. Concurrent binning of machine learning data. US Patent 9,672,474- P. P. Yee Lee et al. Feature processing tradeoff management, WO Patent 2016004062A1- P. P. Yee Lee et al. Interactive interfaces for machine learning model evaluations, WO Patent 2016004075A1

Building backend services for content discovery as well as behavior-based abuse detection at SoundCloud, the world's largest social sound and music platform.Achievements:- Design of a highly available, distributed abuse detection system called Sketchy.- Implementation of the system in Scala based on technologies such as Akka, RabbitMQ, Memcached, Hadoop, and MySQL.- Development of detectors for junk messages (SVM classifier), bulk messages (Jaccard distance estimator), follow spam, and others.- Collaborative filtering-based recommendation engine implemented in C++ based on GraphChi to recommend sounds and creators; data preprocessing in Scala/Scalding/Hadoop.- Audio classification using audio features, sound title/description/tags, and social graph information.- Universal PageRank with handling of dangling nodes, time decay, and heterogenous links.

Research assistant, PhD candidate, and tutor at the Institute of Machine Learning of the University of Potsdam.

As part of my PhD, funded by Strato AG, I designed and implemented a server-sided spam filtering solution running in production at Strato, the second largest European webhoster.Achievements:- Design of a highly scalable discriminative email classifier.- Implementation and integration of a C-library to train the classifier and process/classify emails.- Processing includes one-pass text normalization (MIME parsing, UTF-8 conversation, and HTML/URL decoding) and tokenization, image feature extraction, and classification; all within the SMTP session.- Classifier regularly trained with millions of emails; checks several 100 million emails per daysynchronously while running only on a handful of machines.- Development of management system to monitor, maintain, train and evaluate the spam filter.- Designed and implemented a single-pass GIF/PNG fingerprinting solution in C to detect and block image spam.- Supported the development of an online text clustering solution and regex generator to block/accept bulk messages.

Research assistant and PhD candidate at the Machine Learning Research Group at the Max Planck Institute for Informatics in Saarbrücken.

Research assistant, PhD candidate, and tutor at the Institute of Knowledge Management of the Humboldt-Universität zu Berlin.