“One of the challenges of the machine learning field is simply how fast it moves — there’s always a new model, platform or framework available,” Andrew Lian told Built In Seattle.
For Lian, principal data scientist at Getty Images, the rapid pace of innovation requires a careful evaluation of the available tools. Staying ahead means not just adopting new technologies, but choosing those that bring real value.
“Whether it’s from industry blog posts, conference talks or updates of new offerings via our cloud provider, I ask some of the same questions consistently,” he said. “How much does it cost? How easy is it to implement? Does it give us more flexibility? How long will it take us to see a return?”
This approach helps Lian strike a balance between seizing opportunities for improvement and avoiding unnecessary disruptions to existing workflows.
“Based on those questions, something like a new library from HuggingFace that allows us to trim 10 percent of inference time off of a model for a week of implementation is an easy win,” he explained. “But moving to a new ML platform that would require a vendor agreement is a more complicated decision.”
For Lian and his team, the key is thoughtful decision-making that aligns with Getty Images’ long-term goals. It’s a method that allows them to stay on the cutting edge without sacrificing stability.
Lian shared more about what a day in the life of a machine learning engineer on his team is like.
Getty Images is a global visual content creator and distributor, providing high-quality stock photos, videos and multimedia for creative and editorial use across various industries.
Describe a typical day with Getty Images. What sorts of problems are you working on? What tools or methodologies do you employ to do your job?
Most days, I start the day by syncing with coworkers to understand any emergent problems, developments, or new goals across our broader working group. Over the past several years as our team has grown, my role has evolved into one of enablement for our broader group — how can we standardize, scale and increase our time to market in our AI/ML practice? Our models help us highlight and display our differentiated visual content, create a more diverse search results set and allow us to help our customers search with more natural language. Essentially solving for how we can roll these out faster and to more markets.
I mostly spend my time these days in Terraform, Gitlab CI/CD and AWS to create workflows to support our group. Our goal is to enable our team to train any model with ease and to easily collaborate with our engineering team to move this model through testing, integration and into production. This may involve evaluating new tools, setting up data collection pipelines, or building tools to troubleshoot our ML training jobs.
“Our goal is to enable our team to train any model with ease and to easily collaborate with our engineering team to move this model through testing, integration and into production.”
Share a project you’ve worked on that you’re particularly proud of. What was the process like start to finish, and what impact did this project have on the business?
I recently redesigned the workflow our team uses to train our models. The new workflow allows us to have standard: unit tests, repository structure, code standards and checks for bias. It also allows for integration into some of our engineering tooling and pipelines. Together, these changes increase our time to market, reduce ramp-up for new team members or for members moving to different working groups and increase confidence in our model outputs.
The best part, though, is that thanks to some of the CI/CD capabilities, we can also update our standards and unit tests retroactively on our prior models. If we need to retrain or to update our models, new training runs will need to be compliant with any updated standards.
