Yotam Oren, is the CEO & Cofounder of Mona Labs, a platform that enables enterprises to transform AI initiatives from lab experiments into scalable business operations by truly understanding how ML models behave in real business processes and applications.
Mona automatically analyzes the behavior of your machine learning models across protected data segments and in the context of the business functions, in order to detect potential AI bias. Mona offers the ability to generate complete fairness reports that meet industry standards and regulations, and offer confidence that the AI application is compliant and free of any bias.
What initially attracted you to computer science?
Computer science is a popular career path in my family, so it was always in the back of mind as a viable option. Of course, Israeli culture is very pro-tech. We celebrate innovative technologists and I always had the perception that CS would offer me a runway for growth and achievement.
Despite that, it only became a personal passion when I reached university age. I was not one of those kids who started coding in middle-school. In my youth, I was too busy playing basketball to pay attention to computers. After high school, I spent close to 5 years in the military, in operational/combat leadership roles. So, in a way, I really only started learning about computer science more when I needed to choose an academic major in university. What captured my attention immediately was that computer science combined solving problems and learning a language (or languages). Two things I was particularly interested in. From then on, I was hooked.
From 2006 to 2008 you worked on mapping and navigation for a small startup, what were some of your key takeaways from this era?
My role at Telmap was building a search engine on top of map and location data.
These were the very early days of “big data” in the enterprise. We weren’t even calling it that, but we were acquiring enormous datasets and trying to draw the most impactful and relevant insights to showcase to our end-users.
One of the striking realizations I had was that companies (including us) made use of so little of their data (not to mention publicly available external data). There was so much potential for new insights, better processes and experiences.
The other takeaway was that being able to get more of our data relied, of course, on having better architectures, better infrastructure and so on.
Could you share the genesis story behind Mona Labs?
The three of us, co-founders, have been around data products throughout our careers.
Nemo, the chief technology officer, is my college friend and classmate, and one of the first employees of Google Tel Aviv. He started a product there called Google Trends, which had a lot of advanced analytics and machine learning based on search engine data. Itai, the other co-founder and chief product officer, was on Nemo’s team at Google (and he and I met through Nemo). The two of them were always frustrated that AI-driven systems were left unmonitored after initial development and testing. Despite difficulty in properly testing these systems before production, teams still didn’t know how well their predictive models did over time. Additionally, it seemed that the only time they’d hear any feedback about AI systems was when things went poorly and the development team was called for a “fire drill” to fix catastrophic issues.
Around the same time, I was a consultant at McKinsey & Co, and one of the biggest barriers I saw to AI and Big Data programs scaling in large enterprises was the lack of trust that business stakeholders had in those programs.
The common thread here became clear to Nemo, Itai and myself in conversations. The industry needed the infrastructure to monitor AI/ML systems in production. We came up with the vision to provide this visibility in order to increase the trust of business stakeholders, and to enable AI teams to always have a handle on how their systems are doing and to iterate more efficiently.
And that's when Mona was founded.
What are some of the current issues with lack of AI Transparency?
In many industries, organizations have already spent tens of millions of dollars into their AI programs, and have seen some initial success in the lab and in small scale deployments. But scaling up, achieving broad adoption and getting the business to actually rely on AI has been a massive challenge for almost everyone.
Why is this happening? Well, it starts with the fact that great research does not automatically translate to great products (A customer once told us, “ML models are like cars, the moment they leave the lab, they lose 20% of their value”). Great products have supporting systems. There are tools and processes to ensure that quality is sustained over time, and that issues are caught early and addressed efficiently. Great products also have a continuous feedback loop, they have an improvement cycle and a roadmap. Consequently, great products require deep and constant performance transparency.
When there’s lack of transparency, you end up with:
- Issues that stay hidden for some time and then burst into the surface causing “fire drills”
- Lengthy and manual investigations and mitigations
- An AI program that is not trusted by the business users and sponsors and ultimately fails to scale
What are some of the challenges behind making predictive models transparent and trustworthy?
Transparency is an important factor in achieving trust, of course. Transparency can come in many forms. There’s single prediction transparency which may include displaying the level of confidence to the user, or providing an explanation/rationale for the prediction. Single prediction transparency is mostly aimed at helping the user get comfortable with the prediction. And then, there’s overall transparency which may include information about predictive accuracy, unexpected results, and potential issues. Overall transparency is needed by the AI team.
The most challenging part of overall transparency is detecting issues early, alerting the relevant team member so that they can take corrective action before catastrophes occur.
Why it is challenging to detect issues early:
- Issues often start small and simmer, before eventually bursting into the surface.
- Issues often start due to uncontrollable or external factors, such as data sources.
- There are many ways to “divide the world” and exhaustively looking for issues in small pockets may result in a lot of noise (alert fatigue), at least when this is done in a naive approach.
Another challenging aspect of providing transparency is the sheer proliferation of AI use cases. This is making a one-size fits all approach almost impossible. Every AI use case may include different data structures, different business cycles, different success metrics, and often different technical approaches and even stacks.
So, it’s a monumental task, but transparency is so fundamental to the success of AI programs, so you have to do it.
Could you share some details on the solutions for NLU / NLP Models & Chatbots?
Conversational AI is one of Mona’s core verticals. We are proud to support innovative companies with a wide range of conversational AI use cases, including language models, chatbots and more.
A common factor across these use cases is that the models operate close (and sometimes visibly) to customers, so the risks of inconsistent performance or bad behavior are higher. It becomes so important for conversational AI teams to understand system behavior at a granular level, which is an area of strengths of Mona’s monitoring solution.
What Mona’s solution does that is quite unique is systematically sifting groups of conversations and finding pockets in which the models (or bots) misbehave. This allows conversational AI teams to identify problems early and before customers notice them. This capability is a critical decision driver for conversational AI teams when selecting monitoring solutions.
To sum it up, Mona provides an end-to-end solution for conversational AI monitoring. It starts with ensuring there’s a single source of information for the systems’ behavior over time, and continues with continuous tracking of key performance indicators, and proactive insights about pockets of misbehavior – enabling teams to take preemptive, efficient corrective measures.
Could you offer some details on Mona’s insight engine?
Sure. Let’s begin with the motivation. The objective of the insight engine is to surface anomalies to the users, with just the right amount of contextual information and without creating noise or leading to alert fatigue.
The insight engine is a one-of-a-kind analytical workflow. In this workflow, the engine searches for anomalies in all segments of the data, allowing early detection of issues when they are still “small”, and before they affect the entire dataset and the downstream business KPIs. It then uses a proprietary algorithm to detect the root causes of the anomalies and makes sure every anomaly is alerted on only once so that noise is avoided. Anomaly types supported include: Time series anomalies, drifts, outliers, model degradation and more.
The insight engine is highly customizable via Mona’s intuitive no-code/low-code configuration. The configurability of the engine makes Mona the most flexible solution in the market, covering a wide range of use-cases (e.g., batch and streaming, with/without business feedback / ground truth, across model versions or between train and inference, and more).
Finally, this insight engine is supported by a visualization dashboard, in which insights can be viewed, and a set of investigation tools to enable root cause analysis and further exploration of the contextual information. The insight engine is also fully integrated with a notification engine that enables feeding insights to users' own work environments, including email, collaboration platforms and so on.
On January 31st, Mona unveiled its new AI fairness solution, could you share with us details on what this feature is and why it matters?
AI fairness is about ensuring that algorithms and AI-driven systems in general make unbiased and equitable decisions. Addressing and preventing biases in AI systems is crucial, as they can result in significant real-world consequences. With AI’s rising prominence, the impact on people's daily lives would be visible in more and more places, including automating our driving, detecting diseases more accurately, improving our understanding of the world, and even creating art. If we can’t trust that it’s fair and unbiased, how would we allow it to continue to spread?
One of the major causes of biases in AI is simply the ability of model training data to represent the real world in full. This can stem from historic discrimination, under-representation of certain groups, or even intentional manipulation of data. For instance, a facial recognition system trained on predominantly light-skinned individuals is likely to have a higher error rate in recognizing individuals with darker skin tones. Similarly, a language model trained on text data from a narrow set of sources may develop biases if the data is skewed towards certain world views, on topics such as religion, culture and so on.
Mona’s AI fairness solution gives AI and business teams confidence that their AI is free of biases. In regulated sectors, Mona’s solution can prepare teams for compliance readiness.
Mona’s fairness solution is special because it sits on the Mona platform – a bridge between AI data and models and their real-world implications. Mona looks at all parts of the business process that the AI model serves in production, to correlate between training data, model behavior, and actual real-world outcomes in order to provide the most comprehensive assessment of fairness.
Second, it has a one-of-a-kind analytical engine that allows for flexible segmentation of the data to control relevant parameters. This enables accurate correlations assessments in the right context, avoiding Simpson’s Paradox and providing a deep real “bias score” for any performance metric and on any protected feature.
So, overall I’d say Mona is a foundational element for teams who need to build and scale responsible AI.
What is your vision for the future of AI?
This is a big question.
I think it’s straightforward to predict that AI will continue to grow in use and impact across a variety of industry sectors and facets of people’s lives. However, it’s hard to take seriously a vision that’s detailed and at the same time tries to cover all the use cases and implications of AI in the future. Because nobody really knows enough to paint that picture credibly.
That being said, what we know for sure is that AI will be in the hands of more people and serve more purposes. The need for governance and transparency will therefore increase significantly.
Real visibility into AI and how it works will play two primary roles. First, it’ll help instill trust in people and lift resistance barriers for faster adoption. Second, it will help whoever operates AI ensure that it’s not getting out of hand.
Thank you for the great interview, readers who wish to learn more should visit Mona Labs.
- Decoding Opportunities and Challenges for LLM Agents in Generative AI
- Tailor Introduces ChatGPT Plugin Enabling Conversational Interface for ERP Operations
- Anthropic Introduces Paid Subscription for its AI Chatbot Platform Claude
- Top 10 LLM Vulnerabilities
- AI-Generated Drake Song Submitted for Grammys: A Pivotal Moment for Music and AI