Car Sales Startup Kavak Kicks ML into High Gear with AWS and a Serverless Architecture

Nowadays, you can summon pretty much anything to your house with a credit card and a smartphone. The consumer automobile industry, however, has been slow to find its footing in the digital age. After all, buying a car is a pretty big deal with a lot of pitfalls to be wary of. It almost sounds too good that a car could show up at your door with your name on it, unless there was a way to bypass the rigmarole of the buying experience.

Enter Kavak, the tech startup that’s in the fast lane toward the future of automobile sales. “With Kavak, you can do everything from your house,” says Anders Christiansen, Vice President of Data Science at Kavak. If you’re selling, Kavak will appraise your car based on industry data, pay you for it, and manage the necessary paperwork. Easy as that. “We can come to your house, inspect your car, and give you an offer.” Christiansen explains. If you’re also buying, Kavak will not only recommend models based on your needs, but it will guarantee that your new set of wheels is in tip-top shape, repaired, and subjected to a detailed inspection.

Anders Christiansen, VP of Data Science

In four short years, Kavak has broadened its range of services considerably, now even offering credit finance solutions to buyers through their new program, Kavak Capital. How has the company managed to become a unicorn so quickly while effectively managing and scaling their data operations? According to Christiansen, it all comes down to optimizing the process. Machine learning leveraging a serverless-first infrastructure helped the team get it down to a science.

“We use machine learning in many parts of our business including pricing, recommendations, credit scoring, and process automation. For example, one model estimates the probability that each auto part will need to be replaced, taking into account the car’s age, mileage, brand and other factors,” he says. This information is then used to prioritize which auto parts Kavak inspects prior to purchasing a car, reducing the length of that process by over 70%.”

AWS and serverless

The major game-changer for Christiansen and the team, though, was building it all with serverless services from AWS, like AWS Glue and AWS Lambda. These products allow Kavak’s machine learning engineers to put their own models into production, saving time and manpower.

In the past, engineers would write code for machine learning models and pass it off to an infrastructure team to deploy. “On our team at Kavak, there are a lot of synergies between the different roles in machine learning,” says Christiansen. “Each phase, from defining business requirements, building the machine learning model, and deploying it to production, are intertwined. And who better to monitor the model in the production stage if not for the person who wrote the machine learning code and understands the core business problem being solved? With our serverless architecture from AWS, Kavak’s machine learning engineers and data scientists are empowered with tools that let them easily deploy, monitor, and update their models in production. It is incredibly easy to get something out the door quickly that works.”

A view into the ML process at Kavak

One tool that helps launch new services is an AWS CloudFormation template that includes all the serverless infrastructure for a generic machine learning application using Infrastructure as Code. The template launches services for automatic data ingestion and model retraining with Glue, testing with Lambda and monitoring with Amazon Kinesis Firehose and Amazon CloudWatch. This template makes it easy for a data scientist to copy their Python code they wrote in a notebook and turn it into a reliable and scaleable service for generating predictions.

Another benefit of choosing a serverless infrastructure is that the cost of launching proof of concepts or alternate versions of a model is virtually nothing. This has allowed Kavak to have many different versions of model code and model parameters available at the same time for no additional cost, and only pay for the number of times a service is asked to return a prediction.

Looking ahead, Christiansen believes the next step for his team in their machine learning journey will be digging further into computer vision applications within Kavak’s offering. “We already use computer vision to automate manual processes and extract information about  about a car. But eventually, we imagine using computer vision to identify much more granular information about cars from photographs.” That addition, says Christiansen, would shave even more time off of their workflow.

For now, Christiansen attributes part of Kavak’s rapid success to the efficiency that serverless makes possible. “Serverless architecture and the tools that we are using in AWS has made us much, much more efficient and effective. And helped us build a very talented team.”