The Learning Curve: How Relay Delivery Developed the Platform for an ML-Powered Restaurant Courier Service

Creating a system from the ground up isn’t easy. VP of Data Science Constantine Vitt sat down with Built In to discuss building the data platform that powers machine learning and enables scalable growth at Relay Delivery.

Written by Jenny Lyons-Cunha
Published on Feb. 22, 2023
The Learning Curve: How Relay Delivery Developed the Platform for an ML-Powered Restaurant Courier Service
Relay Delivery
Brand Studio Logo

At the start of every machine learning implementation is something familiar — a learning curve for the humans creating it. 

This is the crux of VP of Data Science Constantine Vitt’s daily work at Relay Delivery, where he is responsible for end-to-end ML and analytics.

Beyond Vitt’s responsibilities to support data infrastructure and deployment processes, he has had a major hand in developing Relay Delivery’s newest data platform. The comprehensive platform will serve as the cornerstone of its emerging ML efforts. 

“We have to build a platform that will process data from restaurants and couriers in a way that becomes consumable by ML models,” Vitt told Built In. “These models, in turn, produce output that is actionable by the business and the rest of the system in real time.” 

 

Relay Delivery team members work in an open office space.
Relay Delivery

 

WHAT RELAY DELIVERY DOES

Inspired by the emergence of “Big Delivery” and consistent hikes in delivery fees, CEO Alex Blum and CTO Mike Chevett founded delivery-tech company Relay Delivery in 2014. Driven by its mission of making delivery more efficient and affordable, Relay Delivery consolidates the backend logistics for restaurants. It offers owners a way to optimize their delivery operations and focus on building their business. “We have a unique business model in the delivery space, paying couriers by the hour rather than per delivery,” Vitt said. “As a result, we must ensure we have adequately provisioned couriers in each market geography and at a neighborhood level.”

 

The need for a new data platform at Relay Delivery was born of its reservation system, which acts as the “primary control lever,” Vitt said, for managing the number of couriers on the platform.

“Without machine learning, a human would have to manually go through our different market geographies, corresponding weather forecasts and other factors to guess how many couriers we ought to provision for the coming days,” Vitt explained.

The launch of the data platform to power ML for its reservation system mitigates human labor for the tedious task with a sleek, automated process. Vitt noted that the automation can be scaled horizontally by incorporating additional processing power in the future. 

As the product owner, manager and principal developer of the data platform and its growing team, Vitt has touched nearly every component of the system and has personally faced each of its pitfalls. 

In the early days of the data platform project, Vitt grappled with the limitless nature of creating something new. 

“One of the biggest challenges when building something from the ground up, is the lack of well-defined bounds and the complete nonexistence of structure,” Vitt said, adding that it has been an immensely rewarding endeavor. 

 

“In the early days of the data platform project, Vitt grappled with the limitless nature of creating something new.”

 

Through trial and error, Vitt has found a state of equilibrium between planning and execution.

“In the beginning, I had a tendency to plan a bit too much and over-engineer things, which would make inevitable pivots along the way more challenging,” Vitt said. “But a lack of planning can quickly result in inefficient code requiring a lot of refactoring.”

 

The Relay Delivery team chats in their office.
Relay Delivery

THE TOOLS TO SUCCEED

To build out the data platform and ML infrastructure, Vitt and his cohorts used a range of tech tools, including:

  • Postgres for real-time data storage
  • S3 and Redshift for long-term data storage
  • Kubernetes for real-time data processing, machine learning model scoring and training 
  • Argo and Airflow for job scheduling and container orchestration
  • Containerized Python applications for the data system
  • Rabbit MQ, a message broker similar to Kafka, for ingesting data in real time.

 

Along with Vitt’s professional evolution, Relay Delivery’s growth has mirrored the development of its ML-enabling data platform. When Vitt joined the startup, he was one of four people on the engineering team, which included the CTO. The data team has grown in tandem with its ML efforts — both data engineers and data scientists have recently joined Relay Delivery’s ranks. 

As Vitt looks into the future, he hopes to welcome more professionals to the team. 

“I firmly believe that having motivated team members begins with hiring the right people,” Vitt said. “As an extension of this, I think it is essential to give team members ownership over their work and align expectations.” 

Reflecting on his start at Relay Delivery, Vitt recalled being drawn to the abundance of compelling challenges — and the thrill of knowing he could tackle them with his expertise and background. 

“I think my team members feel the same way,” he said. “They recognize the opportunity for impact on the business, which in turn creates value for both restaurants and consumers.”

 

Responses have been edited for length and clarity. Images provided by Relay Delivery.

Hiring Now
Zoro
eCommerce • Information Technology • Retail • Industrial
SHARE