How Hooktheory is Infusing Songwriting with AI—Powered by Beam

40% lower latency
by using the RTX 4090

99.99%
Production Uptime

100,000s
of inference requests daily

“Beam’s infrastructure makes productizing our AI straightforward. We can scale up or down when needed, test new GPU options, and keep our users happy. It’s been a really good partnership.”

Dave Carlton
Founder
HookTheory is a music composition platform that is changing the way people learn and compose music. The company started 15 years ago as a project between three friends at UC Berkeley. Today, Hooktheory uses Beam to power their flagship AI model, Aria, which helps users generate melody and chord suggestions real time.

Challenge

A new model leads to a search for a scalable inference provider
Hooktheory’s AI journey began with a collaboration between researchers at Stanford and Google who recognized the value of Hooktheory’s “Theorytab” database. This dataset contains the functional harmonic data of thousands of songs, which is the ideal training set for AI-driven music composition.

When it came time to deploy the trained model, Hooktheory encountered several problems:

• Lack of real-time inference infrastructure. Hooktheory needed GPU hosting for real-time inference at scale—something they hadn’t done before.
• Unable to access elastic compute. Early on, Hooktheory tried using Google’s Vertex AI and AWS Sagemaker and struggled to scale their workloads on-demand. Hooktheory began looking for a serverless approach to scale elastically.
• Reliability and latency. With customers paying for interactive music suggestions, Hooktheory needed constant availability. Any downtime or latency would ruin the creative workflow for its users.

Discovering Beam

Scaling with flexible on-demand compute
Hooktheory began searching for an inference partner and discovered Beam. Several benefits stood out to them:

• GPUs on-demand. Beam’s infrastructure automatically scales to handle real-time traffic and eliminated the need for Hooktheory to manage Kubernetes clusters or autoscaling logic themselves.
• Flexible autoscaling. Hooktheory always wants at least one instance running, with the ability to scale up from there.
• Easy experimentation. With Beam, Hooktheory is able to easily switch between GPU types to find the best balance of performance and cost. They were able to test new hardware by spinning up a Beam deployment, measuring inference time, and finding the best configuration for their model – in this case, an RTX 4090.

Results

Shipping a state-of-the-art model to customers on reliable infrastructure
Powered by Beam, Hooktheory was able to ship Aria to market and introduce the model to their customers. When Beam began offering the RTX 4090, Hooktheory was able to switch from their previous A10G and achieve over 40% faster inference time – a major benefit for their user experience. And thanks to Beam’s infrastructure, Hooktheory rarely encounters downtime for their AI model. “It’s been almost 100% uptime,” said Dave Carlton, Hooktheory’s Co-Founder.

Hooktheory has also benefited from Beam’s customer support – when they needed the ability to eliminate cold boots from their deployment, Beam introduced a new feature to allow a minimum number of always-on instances, ensuring that Hooktheory users never wait for a container to boot.

Hooktheory is focused on enabling more creative and accessible music production. With Beam, Hooktheory was able to ship Aria to production and eliminate DevOps – giving them the ability to focus on building the best possible product for their customers.