Mage launches low-code AI dev tool into general availability

Mage, an archaic term for a magician or someone who makes magic, is now also the name of a Silicon Valley startup that’s demonstrating some magic of its own.

The Santa Clara, California-based company today released to general availability its prize low-code tool for product developers to build AI ranking models. Year-old Mage has been in private beta for the last 12 months working closely with early paying customers to make its tool user-friendly, intuitive, and simple to use, the company said.

After working with hundreds of product developers at Airbnb, CEO and cofounder Tommy Dang saw that those developers knew how AI could be used to improve their product, but that they also had to rely on data science resources to help implement their ideas. Data scientists do not come inexpensively anywhere in the world. 

“People who are working on user-facing features, like engineers and backend engineers – they can code but they didn’t go to school for machine learning or AI,” Dang said. “They definitely know what it is and what is useful, but they don’t have the expertise for that. Existing solutions aren’t designed and built for product developers. So, we provide a web-based tool that empowers those individuals to be able to build AI models – specifically a ranking use case. 

“And we’ve seen that building ranking models in products is much in demand. Let’s say you have a lot of news on your home feed or you have a lot of products you want to sell. People need rankings to optimize that for their users. And usually machine learning and AI are well suited to do that.” 

Use cases include increasing user engagement by ranking articles, posts, comments, etc. on your user’s home feed or increasing conversion by showing the most relevant products for a user to buy, Dang said.

Mage works by first connecting to existing data sources, such as Amplitude or Snowflake. Once a user adds their data, Mage will provide guided suggestions for cleaning and enhancing that data to maximize the model’s performance during training. Once the model completes training, product developers can use its predictions in real-time via API requests, Dang said.

Mage offers a free hobby tier. When a developer or company wants to train larger AI models and use more real-time API predictions, they’ll need to upgrade to the Pro subscription tier.

How the AI is implemented

In order for technologists, data architects, and software developers to learn more about how to utilize AI, VentureBeat asked the following questions of Mage CEO Tommy Dang, who offered our readers these details:

VentureBeat: What AI and ML tools are you using specifically? 

Mage: Scikit-learn, XGBoost, TensorFlow, SHAP.

VentureBeat: Are you using models and algorithms out of a box — from exaPEle, from DataRobot or other sources? 

Mage: When a product developer uses Mage to build a ranking model, Mage will build a unique model for their specific use case. We don’t use out of the box models, we construct them per use case. We use open source algorithms in these models such as linear regression, logistic regression, deep neural networks, XGBoost, etc.

VentureBeat: What cloud service are you using mainly? 

Mage: We primarily use AWS.

VentureBeat: Are you using a lot of the AI workflow tools that come with that cloud? 

Mage: We don’t use many of their AI workflow tools; they didn’t fit our needs and didn’t solve our problems.

VentureBeat: How much do you do yourselves? 

Mage: We use Airflow for data pipeline orchestration and host it on Astronomer. We use Spark to process big data and use AWS EMR to run those Spark jobs. We mostly use AWS for compute. We have proprietary pipelines and workflows for preparing data, training and evaluating models, and serving model predictions.

VentureBeat: How are you labeling data for the ML and AI workflows?  And can you share a ballpark estimate on the amount of data you are processing? 

Mage: We specialize in training on tabular and text data that is already labeled. For data that is unlabeled, we provide guided suggestions that help the product developer programmatically label their structured data.

[And we are processing] billions of data points.