The Critical Role of Data Monitoring in Machine Learning

The effectiveness of machine learning (ML) models is fundamentally linked to the quality of the data they are provided. This holds true not only during the initial training phase, but also after the model has been deployed into a live production environment.

Data Drift and Model Performance

Real-world data is dynamic, constantly evolving with new occurrences. Even seemingly minor alterations in how databases and APIs report or store information can significantly impact a model’s performance. Because ML models typically don’t signal errors when making incorrect predictions, it’s vital for organizations to actively monitor their data pipelines.

Introducing Aporia: An ML Monitoring Platform

Companies like Aporia are addressing this need. The Tel Aviv-based company recently secured $5 million in seed funding to further develop its ML model monitoring platform. The investment round was led by Vertex Ventures and TLV Partners.

The Founder's Journey

Aporia’s co-founder and CEO, Liran Hason, brings a unique background to the challenge. Following five years of service in the Israel Defense Forces, he gained experience as a data scientist at Adallom, a security firm acquired by Microsoft in 2015.

Early Insights at Adallom

Hason later joined Vertex Ventures before establishing Aporia in late 2019. It was during his time at Adallom, however, that he first recognized the difficulties inherent in deploying and maintaining machine learning models.

“I was responsible for the production architecture of the machine learning models,” Hason explained. “This is where I first encountered the challenges of getting models into production and the unexpected issues that arise.”

Aporia's Core Mission

The central goal of Aporia is to simplify the implementation of machine learning models and enable businesses to utilize AI responsibly.

“AI is an incredibly powerful technology,” Hason stated. “However, unlike traditional software, its performance is heavily reliant on data. A key characteristic of AI is its 'silent failure' mode – it doesn’t generate exceptions or errors. This poses a significant challenge, particularly in production, as data scientists have complete control during training.”

The Risk of Unnoticed Data Changes

As Hason pointed out, production systems often depend on data from external sources. If a vendor modifies the data schema without notification, a model – for example, one predicting loan defaults – can become unreliable. The issue may remain undetected for weeks or even months.

Proactive Data Drift Detection

Aporia continuously analyzes the statistical characteristics of incoming data. When significant deviations from the training dataset are detected, the platform alerts users.

Intuitive Monitoring Configuration

Aporia distinguishes itself by offering a visual, IFTTT or Zapier-style interface for configuring monitoring logic. It includes over 50 pre-configured monitor combinations and provides complete transparency into their underlying mechanisms.

From Generic to Customized Solutions

The team initially explored building generic monitoring solutions. However, they realized that this approach would be overly complex and that data scientists possess the specific knowledge needed to define effective monitoring parameters for their models.

Industry Perspective

“Monitoring production workloads is a standard practice in software engineering, and it’s time for machine learning to receive the same level of scrutiny,” said Rona Segev, founding partner at TLV Partners. “Aporia’s team has strong production-engineering experience, making their solution simple, secure, and robust.”