Technical talk | English
Technical talk | English
Theatre 21: Track 5
Wednesday - 15.35 to 16.15 - Technical
Cabify is now the most diverse Mobility as a Service platform in most of the markets where it operates: we support ride hailing with black car and taxis, and asset sharing with motorbikes, kick scooters and electric bikes in over 70 markets across Europe and Latin America. Data & Research team’s mission is to fill up the organisation with more and more automatic decision makers that learn from data generated by our previous decisions, thus making sure every journey we are #oneJouneyWiser.
Our automatic decision makers tackle problems occurring in other businesses, such as fraud control, marketing optimisation or the usage of AI to improve customer support experience but also some problems more unique to mobility such as real time dynamic pricing in a two-sided marketplace and real time assignment of rides to drivers.
In this talk, we will start describing the phases of a Data & Research project within Cabify, and how the different roles play together across each of them. First we do Problem Dimensioning, which starts with more or less anecdotal evidence and finishes when we have rigorously estimated the size of the problem or opportunity. If the dimensioning points to a high impact possible solution, we go to Model Prototyping, where we aim to find the simplest possible model that solves the problem to the extent we are aiming for. This phase typically includes some testing, either against cold data from the archive or real time listening to the marketplace but avoiding to affect a user, so we can take bigger risks. If a viable model is found, next phase is Industrialisation when go full engineering mode to make sure we build a fully monitored, cost efficient, highly scalable, highly reliable solution able to cope with our always growing volumes. At last we have the Inference phase, where we aim to establish a causal relationship between the improvement we are deploying and some measurable experience of our drivers, riders and companies.
Through these steps, we are generating significant amounts of new knowledge for the organisation, and we would like that knowledge to be high quality (i.e. strictly peer reviewed), searchable, reproducible… which we ensure by a process built open source Knowledge Repo project, which Cabify is now an active contributor to. Despite a super-strong bias for the simpler solution, we compete against many of the most AI- savvy companies in the world. Recently, we found that as far as our v2 models were increasing in complexity or we were simply targeting inherently more difficult problems, our aforementioned process had a significant bottleneck in the Industrialisation phase. Thus, we have decided to heavily invest in the building of Lykeion, our machine learning as a service platform. Lykeion empowers data scientists to deploy new models in very demanding production environments (thousands of estimations/decisions per second) without engineering intervention, and also speeds up the process for new models based on the previous ones even in a cross domain fashion, because it is built on top of a centralised feature store.
To finish up, we will showcase Lykeion with some of the more fancy models currently operating on the platform; a route engine based on deep learning that has never seen a map, but rather learnt from our asset trajectories, a smart discounts system that has turned Cabify into the most financially sustainable mobility platform in the world and a real time harassment detection system that help us deliver on our promise of offering a safe ride despite operating in some of the cities were personal safety is more at stake across the globe.