For the past couple of years, we have spent a lot of time working on how to build scalable robust Artificial Intelligence. In the process of doing so, we worked in many different areas, including self-driving maintenance robots, defect detection using drones, medical imaging and drug discovery. These are all examples of areas where robustness and proper estimation of risk are key. The core software we used in these areas is our open-sourced Deep Bayesian Inference engine Borch with 3000+ git commits. Borch itself is a Deep Universal Probabilistic Programming Language written in python which offers the same flexibility in modeling as the autodiff engine it's based on namely "PyTorch". It is this very flexibility that allowed us to tackle these heterogeneous areas of modeling.

In parallel to the work we did on more classical Artificial Intelligence and Machine Learning problems, we also evaluated our solution on various time series models used in finance and demand forecasting. These are models that are particularly difficult to do right and get an accurate estimation of the associated risks. We found that Borch was able to deal with the challenges of long-term effects and short-term effects happening simultaneously by utilizing the inherent ability of neural networks to detect hidden features in data. As it turns out, big data was not as important as deep data and carefully designed inductive biases.

We realized that many of these new modeling mechanisms would be suitable to handle some of the difficulties business science models have had for the past 3 decades. The reason why we thought of this was that I have worked extensively with these types of models since 2009, and they have always required a lot of time and effort to build, preventing many organizations from in-housing them and therefore relying on expensive external consultancies to provide them.