Some games are just completely, utterly, smash-the-controller-and-scream-to-the-heavens unbeatable.
There are quite a few video games out there that are meant to be beaten, but whether it be because of glitches, like in Superman 64, or just poor design, like the impossible-to-clear jump in a ’90s game release of Teenage Mutant Ninja Turtles, some games can’t be beat.
In the “game” of recycling, plastics might be the most difficult boss ever faced.
Very little plastics in the United States end up actually recycled instead of just in a landfill. A Greenpeace report from 2022 found that just 5 percent of plastics are given another life. In 2024, the California Attorney General even filed a lawsuit against ExxonMobil for allegedly deceiving consumers into thinking that far more plastics are recyclable than actually are.
Like an unbeatable video game, there are a lot of steps in the recycling process for plastics that have gone wrong, keeping plastic recycling from leveling up.
Recycling plastics can either be a mechanical or chemical process. Plastics fall into several categories depending on what they’re made of. Most of us know this from the little recycling triangle stamped into the bottom of plastics that surrounds a number ranging from one through seven. Anything stamped as a one, known as a PET plastic, can be recycled chemically. If a number three plastic, like PVC pipes, is anywhere in the mix with other plastics, it contaminates the whole stream (when heated, this plastic can turn into hydrochloric acid).
Aside from the range of plastic materials and the different recycling processes, another challenge is in sorting the plastic. Because the recycling process is so delicate and specific for each type of plastic, the sorting and cleaning process has to be exact. This ends up costing recycling and waste management companies more than they are willing to pay.
But there is hope: One big area of plastic recycling that is starting to get a cheat code — identifying which plastics are which using the power of AI.
Built In Colorado spoke with Hiram Foster, a machine learning infrastructure engineer at AMP, who is designing AI models that can identify what types of plastics are in a batch to help recycling companies make the process more efficient and affordable.
AMP is an AI-powered sortation company that gives waste and recycling leaders the power to reduce labor costs, increase resource recovery and deliver more reliable operations.
Describe a typical day with AMP. What sorts of problems are you working on? What tools or methodologies do you employ to do your job?
At AMP, we’re constantly building and training new AI models to handle the unique material streams our customers process, ensuring we extract as much value as possible. My work directly supports our data team by providing the training data and metrics needed to ensure we deploy the most accurate models for each material. I also collaborate with the modeling team, equipping them with the tools they need to innovate and advance our AI capabilities to the next level.
Each day brings a new challenge, whether it’s refactoring a data pipeline to balance cost and performance, building tools to understand AI model differences or implementing performance checks to ensure our AI doesn’t confuse cleats on conveyor belts with recyclable plastic. When it comes to tools and methodologies, I want to emphasize the importance of customer interviews and having an established design process. These conversations consistently reveal insights that challenge assumptions, clarify goals and keep us aligned with user needs, ensuring we deliver solutions that truly make an impact.
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?
One project I’m especially proud of is the model registry I built, designed to organize over 22,000 trained models, making it easier for users across the company to access key information. When I first arrived, I noticed it was difficult to find details like why a model was trained, its data sources, performance metrics or even who trained it. Though not technically complex, the project required integrating many aspects of our data engine to make this information accessible and actionable.
“One project I’m especially proud of is the model registry I built, designed to organize over 22,000 trained models, making it easier for users across the company to access key information.”
I started by gathering user stories, which led to a build-versus-buy analysis. Ultimately, I built the registry in-house. The registry helps users easily find models, see how they were trained, check how well they performed and track their results on important tasks, like distinguishing between different types of plastic.
Recently, I’ve been working on giving users the ability to fine-tune models for a specific task directly through the registry. It’s a project that keeps giving, helping us optimize our workflows and model training processes long after its initial implementation.
How do you stay updated with the latest advancements in machine learning, and how do you apply them to your work?
Staying updated with the latest advancements in machine learning is a critical part of my role, and I’ve developed a few methods that help me keep pace with the rapid developments in the field. The biggest piece of advice I have is trying to figure out how to incorporate it into a natural part of your daily work.
One of my favorite ways to stay updated is through our journal clubs at AMP. The modeling team take turns sharing interesting journal articles and comparing them to our own practices. These discussions are incredibly valuable, and we’ve seen several concepts from these sessions make it onto our roadmap to experiment with.
As much as I can, I try to offload the job of staying informed by simply curating my newsfeed algorithms. Google, Feedly and different newsletters surface relevant news, research papers and tools that I can check out when I have the urge to doom-scroll.
Finally, whenever we need to evaluate solutions for a particular challenge, I use it as an opportunity to see what's out there, read up on the latest tools and frameworks, and skim through sites like Reddit to gauge user experiences and get a sense of how people are tackling similar challenges.
