AI observability startup Aporia nabs $5M to provide guardrails for AI

Machine learning observability startup Aporia today emerged from stealth with $5 million in venture capital funding. The company says the proceeds will support the unveiling of its platform for AI models, which enables companies to monitor AI running in cloud environments.

Businesses are investing upwards of $50 billion annually on AI adoption, a recent report from the MIT Sloan School of Management and Boston Consulting Group found. But a lack of ability to detect issues in models as they enter production could be undermining investments.

“AI needs guardrails,” Aporia CEO Liran Hason told VentureBeat via email. “Companies need to have confidence in their machine learning models, and the only way to get there is by robust monitoring to ensure they’re doing what they’re supposed to do.”

Hason, a veteran of the Israel Defense Forces’ elite 81 intelligence unit, was one of the first employees at cloud security company Adallom, which Microsoft acquired for $320 million in 2015. At Adallom, he led the machine learning production architecture, which served as many as millions of users.

“I founded the company in late 2019 after leading machine learning production architecture at Adallom and then working at Vertex Ventures VC where I was involved in dozens of startup investments,” Hason told VentureBeat via email. “It seemed natural to use my development best practices, hard-learned lessons with data science challenges, and my gravitation towards startups to start a company that would apply best practices from production engineering and adjust them to machine learning, in the hope of transforming doubt into trust and build what analysts often call ‘responsible AI.’”

Aporia lets data scientists create, maintain, or modify monitors for models and set alerts that trigger notifications via email, Slack, and other channels.  The Aporia platform can be installed with a few lines of code and set to monitor billions of daily model predictions asynchronously. Alongside its public cloud offering, Aporia provides a managed on-premises solution for enterprises with data privacy and security requirements.

Machine learning models can work perfectly in the experimentation phase but start to drift in production over time due to changes in their datasets, Hason explained. Something as routine as a company expanding into a new market can affect the performance of a model. Customers and businesses typically suffer the consequences — predictions based on the wrong data are flawed, resulting in unintended outcomes and in turn lost revenue.

“Companies are struggling to keep watch of their AI in the ways that matter for their specific machine learning model and use case,” Hason added. “Aporia’s platform contains three pillars: (1) visibility, allowing data scientists to explore production data easily, (2) monitoring, the beating heart of the system, where users can implement any monitoring logic they’d like and adapt it to their use case and investigation, and (3) toolbox, for root cause analysis. Aporia aims to be the place where organizations manage the reliability of their models, and ensure responsible usage, whether in regard to performance or bias and fairness matters.”

Sixteen-employee Aporia has rivals in data reliability startup Monte Carlo and WhyLabs, a startup developing a solution for model monitoring and troubleshooting. There’s also Domino Data Lab, a company that claims to prevent AI models from mistakenly exhibiting bias or degrading.

But according to Hason, Aporia’s differentiator is its experienced team. Already, the company’s platform is being used by roughly a dozen users across over 11 “multi-billion-dollar” companies. One organization is tapping Aporia to monitor a model that predicts whether an applicant will be able to repay a loan without defaulting.

“We had one case where the credit history data provider had changed the schema of the data without notifying anyone, resulting in a significant drift in model’s behavior, leading it to approve or deny loans unjustifiably,” Hason explained. “Without a proper monitoring system in place, it would only have been discovered a few months later once loans were starting to be defaulted on and there was major revenue loss. However, with Aporia, they got an alert about that drift on the very same day the problem had started, which allowed them to react quickly and avoid potential deficiencies.”

When asked about Aporia’s fundraising, Rona Segev, managing partner at investor TLV Partners, said, “Monitoring production workloads is a well-established software engineering practice, and it’s past time for machine learning to be monitored at the same level. Aporia’s team has strong production-engineering experience, which makes their solution stand out as simple, secure, and robust.”

Vertex Ventures and TLV Partners led Tel Aviv, Israel-based Aporia’s seed round.