SigOpt Case Study: Carbon Relay

About Carbon Relay

Carbon Relay is a New York & Washington DC-based AI startup, leveraging deep learning, speech recognition, and computer vision to deliver autonomous, interactive avatars and personas. Earlier this year, Carbon Relay launched mai Social, available on the App Store for iOS devices. mai is a social network for your personal avatar. Users create a digital avatar and customize their voice, style, and animations in real-time.

The Challenge

Being a startup, for Carbon Relay, time and resources are precious commodities. The problem they are trying to solve is complex and requires the use of different architectures (CNNs, RNNs, etc.) for various processes (NLP, image detection, etc.). Having several PhDs who bring expertise on Deep Learning on their engineering team, they still found themselves spending months of their precious Ph.D. time tuning the architecture and the hyperparameters of their neural networks to get the optimal performance their users demand.

Why SigOpt

According to Dr. John Platt, “We knew that the traditional approaches — grid search, Monte Carlo search, simulated annealing or even population-based algorithms — wouldn’t scale well to our problems, while also delivering the optimal performance we required.” Carbon Relay explored open source Bayesian optimizers, such as HyperOpt and MOE, and found that none had all of the functionality they wanted (parallelism, multimetric optimization, etc.). In the course of their evaluation, they found that not only did SigOpt beat their experts at optimization, but it was able to do so much faster, with little to no monitoring or intervention required. As well, they found that SigOpt was responsive to both their questions and feedback.

The Benefits

The primary benefit of using SigOpt for Carbon Relay is that they no longer have to worry about hyperparameter optimization.

“We can keep our experts focused on the tasks core to our business, and entrust the SigOpt platform to find the optimal hyperparameter configurations for our models, irrespective of the data type and model type. Every morning, our engineers check the current state of the optimization experiments and deploy the best models as configured by SigOpt” said Dr. Platt.

SigOpt relies on a number of AWS services to make it possible to support our customers, like Carbon Relay. SigOpt uses EC2 compute for our customers’ experiments, VPCs to secure any sensitive customer data, and RDS for long-term storage of results so that customers can analyze results and maintain the history of our optimization experiments. Perhaps most importantly, AWS provides SigOpt the ability to scale these resources up and down as our customer base grows, and customer usage changes, allowing SigOpt to maximize the efficiency of our infrastructure spending.

Carbon Relay has been getting value from SigOpt since early 2017.

Next Steps

Visit to learn more, or email [email protected] to talk to our sales team!