Josh Elliot, Head of Operations at Modzy talks about the significance of an open architecture solution in deploying Enterprise AI at scale.
1. Tell us about your role at Modzy.
As Modzy’s™ Head of Operations, I have the privilege of wearing a lot of hats from driving Modzy’s long-term strategy for business and product to ensuring aggressive and successful growth of the business to managing all hands-on operational aspects of the company. Plus, I get to work with a great team of people responsible for building an innovative product!
2. What are the applications or rather opportunities you seek to have with your product?
Modzy helps organizations move AI from the lab into production systems, whether you’re a small team or aiming for enterprise scale. There’s currently a huge gap between the process of handing off AI models from data science teams to software development teams building AI-powered applications creating missed value and opportunities for organizations looking to work smarter, faster, and more efficiently.
Modzy makes it easy to turn any application into one that’s powered by AI. It is the foundation for everything needed to manage your AI in the long term – centralized deployment, management, monitoring, governance, and security for all AI models in use in production systems. Organizations such as financial services, energy and utilities, life sciences, and federal government agencies, to name a few, can use Modzy to reduce the risks associated with deploying AI in their enterprise systems.
3. How did you define the vision of Modzy?
Modzy was built with a distinct purpose: speed AI adoption in pursuit of mission and business applications where it matters most. After years of working with customers, talking to venture capital firms, and academics focused on machine learning and AI, and painfully watching customers labor through every single model build, we were in the position to help them step back and realize there was a better, more efficient way. Modzy solves last mile challenges with deploying, managing, monitoring, and securing AI models in production systems, allowing teams to start seeing the real value of AI. Since our public launch in November 2019, we’ve grown from a small team of four to a team of many passionate about making AI real across an organization.
4. What are some of the unique lessons you have learnt from analysing your customer behaviour?
First, there’s really a lot of great experimentation happening across industry with some sectors more mature than others. With that perspective, here are several patterns observed:
1. The real-world is messy—From dynamic problem sets and rigid software architectures to inconsistent or even constrained IT and networking environments, lab simulations are difficult. Therefore, only about half of AI experiments out there make it to production.
2. The collaboration and hand-off from data scientist to software developer is not where it needs to be. Too many organizations don’t have the processes, tools, and technology in place to iterate and deploy new AI capabilities quickly. That’s why we’re seeing it take on average between 5 to 9 months to deploy a single model.
3. Unfortunately, I think organizations have been taking a laissez-faire approach to AI governance and are missing an opportunity to provide guidelines that promote explainability, transparency, accountability, and security – and not without mention, a stance that promotes the development of ethical AI.
4. There’s a build/buy debate happening inside many organization’s procurement teams with special consideration given to the cost associated with in-sourcing high-quality data scientists, preparing data sets, and the time and infrastructure resources required to design, train, evaluate, deploy, monitor, and maintain a bespoke model.
5. What are some of the distinctive features of Modzy’s platform for Enterprise AI?
LAST MILE DELIVERY CHALLENGES IN DEPLOYING AI FOR ENTERPRISE SCALE REQUIRE TOOLS THAT ADDRESS THE HARDEST PART: DEPLOYMENT, MANAGEMENT, MONITORING, AND SECURING AI MODELS FOR SCALE.
Modzy is an open architecture solution that simply integrates with your existing model training tools and tech stack, allowing you to turn any tool or application into one powered by AI with little disruption to existing operations.
With click-button integrations for the most popular model training tools and our powerful APIs and SDKs, Modzy removes the friction during the handoff between data science and development teams—making it painless to build an AI pipeline. Rather than lock you into a specific tool, Modzy allows you to connect into new systems or training tools to add or update functionality as needed. Additionally, Modzy’s patent-pending adversarial defense and explainability solutions are charting the course for a new level of AI performance. By better understanding how AI models work and behave, we can better protect, secure, and explain how they work, ensuring trustworthy, defense-grade AI in use for production systems.
These capabilities are supplemented by the expansive capabilities available in the Modzy model repository (for storing your own models) and marketplace, which offers a more diverse and niche collection of hard-to-find models not found anywhere else that are already vetted, pre-trained and retrainable.
6. Modzy was recently in the news for partnering with DarwinAI. Can you elaborate more on the same
The Modzy model marketplace has 100+ trusted, pre-trained AI models from leading machine learning companies across a wide range of AI capabilities. Models in the Modzy marketplace can be retrained on customer data. The marketplace also serves as a repository for any models (including customer models) that are deployed and managed via the Modzy platform.
Recently, we’ve added several partners to the marketplace, such as Kensho, DarwinAI, and TrueFace, which offers NIST-certified solutions. Our partners demonstrate our commitment to furthering the development of trustworthy AI for critical applications. The addition of partners such as DarwinAI expands our offerings in the life sciences space, especially at a time when AI innovation is so crucial in the fight against the COVID-19 pandemic.
Modzy aims for maximum transparency of all models listed on the Modzy marketplace to help our customers make informed decisions about which machine learning models will meet their needs. Before a model is listed in the marketplace, we conduct a rigorous due diligence review to evaluate the model provider as well as other internal policies that could lead to potential risk. When a partner distribution agreement is signed, model contributors must also meet the transparency standards of the Modzy model marketplace and endeavor to adhere to our community guidelines. When submitting models, we request each machine learning model be accompanied by comprehensive documentation about the model’s training data, architecture, training approach, testing and validation strategy, performance metrics, inputs and outputs, inherited licenses, and any relevant research referenced by the model designer. We also ask that the model designer and the senior-most engineer sign a testimony confirming the accuracy of all the information provided. Once models are deployed to Modzy, model performance is monitored, and model enhancements are made by providers when necessary.
7. What are some of the common pain points that your customers commonly approach you with?
Customers approach us with a myriad of problems, however they all bubble up to how you make AI real for an enterprise. Maybe they’ve already invested in AI training tools, but struggle with productionizing the models. Gartner Research shows that it takes teams 9 months to deploy AI models, and the handoff from data science to development teams is fraught with unnecessary complications. IT organizations tasked with overseeing AI implementations lack insight into AI usage or performance across the enterprise, and have no semblance of risk management for their AI programs. We’re approached by many of these stakeholders who recognize the need for a better way to deploy and manage AI models. Through detailed model documentation, model management, monitoring, drift detection, governance, and comprehensive security— including adversarial defense—Modzy provides all of the necessary components of a full ModelOps solution.
8. What advice would you like to give to the upcoming AI-based tech start-ups?
There are three things I would suggest:
1. Don’t underestimate the importance of a thoroughly constructed go-to-market strategy. That means understanding the market size and demand for the problem you’re trying to solve, talking to customers to confirm pain points and test your value proposition, knowing your competition, and designing a business model that doesn’t put your product into your customer’s value stream.
2. IF YOU’RE GOING TO DESCRIBE YOUR COMPANY AS AN AI TECH STARTUP, PLEASE MAKE SURE THE “EMPEROR IS WEARING CLOTHES”, I.E., YOUR PRODUCT EMBEDS OR OFFERS AI SOLUTIONS.
Taking advantage of a still-maturing AI buyer will only slow market adoption.
3. Consider making your IP available to a broader customer base by deploying your AI models to the Modzy marketplace. J
9. Can you give us a sneak peek into some of the upcoming product upgrades that your customers can look forward to?
We’re working on some improved workflows for data scientists to speed up model deployment, in addition to more integrations for analytical tools, business applications, and more. We’re also working on some optimizations and accelerations on GPUs, as well as exploring some exciting innovations at the intersection of AI, 5G, and edge computing.
10. Which is the one AI breakthrough you will be on the lookout for in the upcoming year?
We’re going to see a huge shift in the risk management posture for AI deployments, likely due in part to a major security incident that sheds light on the current lackadaisical approach to AI risk management. A proactive risk management posture for managing AI starts with building the core foundation and repeatable process for deploying, managing, monitoring, governing, and securing AI models which enables you to then put the same rigor in place with managing AI systems and applications as you would with any other IT implementation.
11. What is the one leadership motto you live by?
There are two quotes I really admire and influence my leadership style.
“If your actions inspire others to dream more, learn more, do more, and become more, you are a leader.” John Q. Adams
“There are basically two types of people. People who accomplish things, and people who claim to have accomplished things. The first group is less crowded.” Mark Twain