References

Below you will find a selection of our past projects.

SoundCloud

SoundCloud is one of the leading music streaming companies worldwide. It is mentioned along Companies such as Spotify and YouTube.

For SoundCloud we developed recommender systems, spam detection systems and optimized their search ranking. We designed the systems from scratch, implemented and maintained them.

From big data processing to model training to prototyping and online A/B testing. The solutions had to scale up to hundreds of millions of users. The models were trained on billions of data points.

We were additionally involved in building up SoundCloud’s data organisation and had an essential influence on SoundCloud’s success.

We published a paper which describes one of the implemented models.

case-study-image

BSDEX (Stock Exchange Stuttgart)

BSDEX is a crypto exchange platform and a subsidiary of the German stock exchange Stuttgart.

BSDEX wanted to become a data driven organization and asked us to assist them.

In order to do that we introduced new technologies and moved parts of the old infrastructure into the AWS cloud.

Technology is just one part in order to become a data driven organization. More importantly is to enable the people. For that sound processes and an appropriate organizational structure needs to be in place. We decided to organize the company around a data mesh and defined necessary data processes.

Technology and enabled people helped BSDEX to get the most value out of their data.

case-study-image

Deep Breath Intelligence

Deep Breath Intelligence (DBI) combines high-resolution mass spectrometry with advanced machine learning models to provide breath metabolic profiling. This allows for instance to predict diseases in real time based on the human breath.

For DBI we ported their existing Matlab models into a modern development environment. Among other things this included retraining the models with neural networks and deploying them into a production ready scalable cloud environment on GCP.

We also worked on improving the algorithms to extract features from Mass Spectrometry data. This allowed us to speed up the training process massively which in turn reduced the cycle to build models for new disease use cases.

case-study-image

Idagio

Idagio is the leading music streaming service for classical music.

For Idagio we designed and implemented a music recommender system from scratch. The final model is a deep neural network and was based on YouTube’s production recommender system.

The model is using various input signals such as metadata information, collaborative filtering signals as well as the raw audio data itself.

The system had to scale up to millions of audio tracks.

case-study-image

Muru Music

For Muru Music we built a system to classify audio tracks into their main and sub genres.

We used a deep convolutional neural network which operated directly on the raw audio features. No high level features were used.

The system was trained on multiple large scale production catalogues from BMG, Sony and Warner. A total of 2 million tracks were used to train the model.

In order to deal with these data set sizes we built a data platform on top of Google Cloud Services. All pipeline steps were fully automated: preprocessing of the audio, model training, model prediction and rich evaluation. This data platform enabled us to quickly run experiments and iterate on the model.

Google theirselves referenced this solution in an article.

case-study-image