Machine learning continues to be applied to use cases across industries. Ksenia Khlyustina discusses her experience in software development and machine learning, and how FLYR Labs is using machine learning and AI to predict demand and optimize pricing for airlines.

Speakers: Ksenia Khlyustina, FLYR Machine Learning Platform Manager, and Mikhail Chrestkha, Google Outbound Product Manager

Date Presented: April 2, 2022

Duration: 16:31

01:34 – Machine Learning Use Case
FLYR’s sole use case for now is taking all historical data, competitor data, data from future events and origin and destination of a flight, and putting that data into models to output demand forecasts. These forecasts are used to optimize pricing for airlines, who use it in their production systems.

02:12 – The FLYR Data Science Team
Because FLYR is built around data science, we have a data science organization that consists of three teams – the science team, the machine learning platform team, and the machine learning infrastructure team – as well as a separate validation team.

The data science team is responsible for building the model for the platform, and the machine learning platform team then brings that model to life. The machine learning infrastructure team focuses on the tools the other teams need to design and build the platform. The validation team then tests the models to see how they are performing in real life.

04:35 – Architecting on Google Cloud
FLYR’s had several reasons for choosing the Google Cloud. The first was Google’s startup powering program, Google Bootcamp, which allowed us to use the cloud platform free of charge to start up. The second reason was Google’s support for BitQuery and Kubernetes, which is why FLYR remains with Google today. Currently, FLYR’s machine learning platform uses other open-source GCP services like DataFlow, Vert CI, Cube Flow, and Tensor Flow.

06:42 – Components of the Machine Learning Architecture
There are three main pipelines, or components, to FLYR’s Cirrus platform – the training dataset generation, model training, and daily in.

07:15 – Training Dataset Generation Pipeline
The Training Dataset Generation Pipeline demands the most resources. This pipeline uses cube flow as an orchestrator, pulling several years of airline historical data from BitQuery and running it on dataflow using in-beam programming model as an output. The pipeline generates TensorFlow records that are put into Google Cloud storage and registered in the model metadata store. Although this pipeline runs on Google TPU, it still takes hours to generate this training dataset.

08:20 – Training Pipeline
The Training Pipeline takes the records from the previous pipeline and a model source code and generates a trained model in a format of a TensorFlow saved model as the output. FLYR puts this in Google Cloud Platform and Google Cloud Storage and registers it in our internal model metadata storage tool. FLYR also keeps a history of the datasets so they can be reused later.

10:25 – Daily Inference Pipeline
This pipeline listens to client data, collects nest server data, and uses Argo workflow to run data flow and pulling model files from Google Cloud Platform using model data storage. This provides predictions in the form, which the platform can deliver to the customer in whatever form they need – API or text files. These predictions also help FLYR build metrics which empower our user interface and give analysts a place to adjust their model inputs and pricing decisions.

12:16 – Designed to Scale
In order to scale, FLYR’s goal is to have one customizable and configurable platform that serves all airlines. This allows us to have one product team and core engineering team that works on the same platform all the time, while the customer success team takes the customer data and transforms it into the canonical data model to provide the outcomes the customer needs.

FLYR also plans to assist customers beyond economy fares by expanding into premium fares, ancillaries, cargo management, marketing, and more. We’ve designed the pipeline to allow us to add as many models necessary and run them in parallel.

14:35 – Future Enhancements and Conclusion
FLYR is currently planning future enhancements to its platform tooling, including a feature store to version and better manage features. Because the Cirrus platform generate completely new data daily, we wouldn’t need to regenerate an entirely new history to be able to run a new inference. Another plan is in place for dynamic pricing, to receive a request for pricing from a customer and be able to respond immediately with an optimal price.