With the increased use of artificial intelligence and machine learning models across various domains, it has become pertinent to ask if the predictions these models make are correct and justified. Censius, an AI observability platform, seeks to address these aspects of ML models. Although the company is headquartered in Texas, a large part of the team works out of India. 

In an exclusive interaction with Analytics India Magazine, Ayush Patel, founder of Censius, speaks in detail about the company’s operations, business model, tech stack and future plans. 

AIM: What is the story behind Censius?

Ayush: We started off trying to do privacy-preserving machine learning for healthcare that could be used to learn from everyone’s data without sharing or collecting it in one place. We spent six months at it and realised that the healthcare industry in the US is strictly regulated – it’s hard to get something moving forward. But while doing this, we also realised that everyone is learning how to clean and organise data, and design and build models. Still, there’s hardly any learning on how to take what you’ve built and make sure it’s accessible and usable by your end users. Also, the constant cycle of learning from new data, taking it into production, and improving this was not out there both from the tooling and knowledge perspective. Knowledge was siloed in some of the top few organisations. Thus, Censius came up.

AIM: What does Censius precisely do?

Ayush: We are in the business of empowering data science and machine learning teams. They know what needs to be done but don’t have the resources and tooling to enable that, so we supercharge them. If a website fails, you get an error message “404 page cannot be found”. However, when models fail, you don’t see that. Over a timeframe, models start degrading in quality. We help them understand using our platform if the model is failing. We essentially help these teams identify where the model is underperforming, potential issues, and how they can be improved. 

AIM: How do you help companies identify if their models are failing?

Ayush: Our platform is a B2B SaaS platform. Our clients lock data that is going into and coming out of their models on our platform and can get all these insights on monitoring, analysing and explaining things. After the model features are locked, from there on, everything they do is on a visual and no-code interface where they see all these different insights. We create something called monitors that are basically thresholds. And if anything fails, our clients get alerts on their preferred medium of communication. Thus, if there’s an issue, monitors are triggered and our clients can analyse it and create, drag and drop custom dashboards and charts to dive deeper to identify the root cause in the entire pipeline. 

AIM: What are the tools that Censius AI uses for data modelling, orchestration and model monitoring?

Ayush: We are the tool – an infrastructure tool for ML and data science teams. An ML infrastructure stack involves a lot of different components that AI teams have to put together to give them solutions. Similarly, they use a combination of different deployment tooling. Post-deployment, they need a monitoring solution, which is what Censius is. We are one of the tools that AI companies will use in the upcoming timeframe.

AIM: What does Censius’ tech stack comprise of?

Ayush: Our tech stack is really broad. It is built across different languages, at least the popular ones. It is cloud native and runs independently on Kubernetes so one can take it and deploy the software into any cloud provider or on-prem on a computer. That is something that many newer companies, especially Platform as a Service (PaaS) providers, adopt.

AIM: By what means is Censius trying to ensure explainability in AI models?

Ayush: We’re making tooling available that lets companies connect their models or data and get insights. There have been a lot of open-source approaches built around explainability, but they are not accessible. Yes, one can find some explainable approaches in some research papers, but how does it go from there to the product manager or business analyst so they can understand why a model made a decision for an end client. So we make it accessible on our platform. We are also working on many different features, which we will roll out in the upcoming months. These are mostly around bias tracking and fairness analysis of the AI models. 

AIM: How does Censius ensure that AI is responsible enough?

Ayush: Responsible AI is a broad term – very domain specific. These are super early days for it. Responsible AI demands intense effort from domain experts to understand what exactly it means for a model to be biased. Models inherit biases because, as humans, we bring in a lot of bias when we look at things, and models reflect that. So to enable responsibility, I think one of the steps is to make the people who build models aware of what’s going on and also, once your models have been built, monitor how these are working. 
In terms of how we approach it, we do a bunch of work around helping these experts at their company understand what it means to be biased and how they can be more responsible. One part is coaching and consulting, and the second is around using the right tooling.

AIM: You are working in a very niche area, do you have competition?

Ayush: We are among a handful of companies that started roughly around the same time and are working towards a similar goal. It is a super niche domain where even the larger cloud providers don’t have the same level of things we are providing, and they are still figuring it out. So in that capacity, we are one of the early startups to dive into the space and build upon it.

AIM: Who are your clients?

Ayush: We have a wide spectrum of clients. Typically, there are two categories of clients that engage with us. First, mid to later-stage startups using AI as their core, like AI-powered insurance companies. The second category includes larger enterprises with large data teams using AI initiatives across the organisation.

AIM: Do you have plans to expand in terms of domain and geography?

Ayush: We are a team of 22 people right now and are increasingly expanding the team. We are actively hiring across all verticals – data science, product marketing, sales, etc. Regarding geography, we’re slowly expanding and targeting more European countries because AI regulation is happening there. It is clearly mentioned in their AI Act that most AI solutions in the market fall under the high-risk category, and there’s a need to adopt post-deployment monitoring solutions for them– exactly what Censius does.

AIM: A large portion of your team is based in India. What is India’s contribution to your business?

Ayush: Many US corporations have large portions of their ML teams residing in India. So even though the companies may not be Indian, the ML talent is from India, and they are the ones making decisions. Also, many homegrown startups like e-commerce and delivery companies are growing in leaps and bounds using AI. So there are a lot of advancements happening in India when it comes to machine learning.

AIM: What, according to you, is the future of MLops?

Ayush: In the initial days, getting a website live and making it accessible to the user was a big ordeal. Today, an entire website can be designed and deployed in a matter of days or even hours. The situation is similar for machine learning currently. It is a big task to collect data, build models and monitor them. It requires great collaboration across machine learning engineers, back-end engineers, data scientists, domain experts and product managers. So, our vision is to take it to the level that websites have reached now and make it accessible to every organisation. However, it needs to be done in a responsible way.