Machine learning observability startup Aporia Technologies Ltd. is adding to its arsenal of tools that provide teams with greater visibility into how their artificial intelligence models work.

The company announced the availability of its new Production Investigation Room today. It’s described as a root cause analysis tool for large language, natural language processing and computer vision models. Aimed at data scientists, machine learning engineers and analysts, it provides an easy-to-use digital environment for those users to carry out real-time data analysis of their AI models, according to the company.

Aporia, which raised $25 million in a Series B funding round in February 2022, is the creator of a comprehensive observability platform for AI teams. Its platform enables teams to create customizable monitors that can detect problems in AI models, such as biased predictions, and “hallucinations,” which is when AI starts fabricating responses. With Aporia, teams can even detect if the performance of their AI models degrades over time.

With the launch of Aporia’s Production IR tool, teams now have a better way to understand the cause of any problems that occur in their AI models, and from that they can come up with ways to fix those issues.

Aporia explains that investigating production data to get to the bottom of any issues affecting AI’s reliability is a complex and time-consuming process, hampered by limited collaboration and code changes. With Production IR, the investigative process can be streamlined, the company said, with no need for any coding skills. It provides a simple user interface to navigate through an AI model’s production data and find insights that can aid in root cause analysis investigations.

Production IR provides a number of features to aid effective investigation, the company said, with segment analysis, data statistics, drift analysis, distribution analysis and incident response tools all included in its armory. With its incident response tool, decision makers can address any potential challenges or issues with their models as soon as they appear.

The tool comes with an embedding projector capability that enables users to visualize unstructured data in both 2D and 3D using a technique known as Uniform Manifold Approximation and Projection. According to Aporia, this can help teams to identify different clusters within a dataset and any patterns within them that might be responsible for model performance issues.

According to Andy Thurai, vice president and principal analyst at Constellation Research Inc., many of the new features introduced in Aporia’s new tool can be found in rival MLOps platforms already. “However, the incident response feature, alongside segment analysis, data statistics, drift analysis and distribution analysis, are good additions to its toolbox,” he said. “The 2D or 3D visualization of unstructured data is an interesting concept too. It can potentially allow users to identify patterns for each cluster when analyzing the production data.”

Aporia Chief Executive Liran Hason promised that Production IR will be a game-changer for teams investigating machine learning events and anomalies. “Data scientists and engineers now have a fast and effortless way to extract valuable information from their production code,” he said. “Our goal is to enable an innovative and effective root cause analysis process, so users can swiftly understand the factors impacting their model’s performance.”

Image: Aporia