In this interview, we speak with Georgios Ilias from Novo Nordisk! Georgios is a Senior AI Engineer at Novo Nordisk, and has over a decade of experience in Natural Language Processing and Deep Learning. At Novo Nordisk, he has been instrumental in implementing generative AI solutions, which have led to significant savings and advancements in clinical data access through the FounData initiative.
In his upcoming NDSML Summit 2024 presentation, Georgios will share insights on agentic architectures, highlighting their role in creating stable, scalable AI systems. He emphasizes that these architectures are essential in healthcare, where sensitive and complex data require specialized approaches for robust AI applications.
Hyperight: Can you tell us more about yourself and your organization? What are your professional background and current working focus?
Georgios Ilias: I am a senior AI engineer at the AI Center of Excellence of Novo Nordisk. Novo Nordisk is a pharmaceutical company, with a focus on diabetes, obesity, and rare diseases. I have over a decade of experience in Pattern Recognition and Natural Language Processing. The AI CoE is a centralized team. It works across the enterprise to rapidly implement AI solutions across the value chain. I am one of the founding members of the AI Lab, an internal team that facilitates generative AI adoption through thought leadership, internal tools, and strategic partnerships. We have been quite successful, managing to be amongst the earliest companies and teams to take advantage of the technology.
Our internal tools generate more than 300 million in savings yearly. Perhaps more importantly, we have enabled teams across Novo Nordisk, including in Development and R&ED, to create effective and robust solutions to core problems using generative AI, often utilizing extremely sensitive and mission-critical data. I am directly supporting one such strategic use case – FounData, a team with a mission to democratize access to clinical data and insights from that data.
Hyperight: During the NDSML Summit 2024, you will present on agentic architectures and dividing AI to conquer stability. What can the delegates at the event expect from your presentation?
Georgios Ilias: At Novo Nordisk, we had early and significant success with Generative AI adoption and application. That journey brought lessons. One such lesson is how ill-equipped we and the industry in general were, and largely still are, at harnessing the power of that technology and mitigating its weaknesses. Generative AI can directly and effectively solve certain types of shallow, text-centric use cases, cases we internally call Citizen AI. Summarization, writing emails, and taking notes can all be solved or streamlined with little consideration for best practices or AI-centric architecture. This applies to almost anything with tolerance to variance that is closely related to language generation. That is not true for more strategic and specialized use cases. Those allergic to variance, where stability, control, and determinism are crucial, data is sensitive and complex, or project scalability and extensibility are key. Those are radically different and far more demanding projects. If the former is a tiny hill in Denmark, the latter is Mount Everest in comparison, and it makes sense we wouldn’t use the same equipment or strategy to conquer both. At Novo Nordisk, we have made great progress in building such equipment: tools, and foundations that can support such cases.
At the summit, I will share some of those tools. Agents, what they are, architectures that utilize them, and a framework that implements those architectures to add stability and control back to the generative AI tech stack. Furthermore, I will show how we’ve leveraged this framework to create robust systems to utilize some of the most sensitive and complicated types of data in healthcare – clinical trials.
Hyperight: What inspired you to focus on agentic architectures? How do you see healthcare evolving in that context?
Georgios Ilias: The inspiration mostly came from previous experience and success I had with solving complicated tasks using multi-agent systems, in the context of reinforcement learning and evolutionary algorithms. We would often look at optimizing each agent for a single type of task. We would then rely on agentic collaboration to solve more complicated super tasks. This approach of simplification and collaboration has proven to be effective when applied to generative AI.
As for your second question, I see agents as a critical component and a natural next step to the practical utilization of AI. Even more so in healthcare, where data is both complex and incredibly sensitive. And failure is not an option given our mandate in society. Internally, more and more teams reach the same conclusions and reach out to adopt our framework and architecture. In the broader industry, similar patterns emerge, with recent comments by the CEOs of both OpenAI and Nvidia on agents seemingly confirming that thesis. I see work with AI evolving into teams making specialist agents and combining their capabilities to solve tasks of increasing complexity. I see teams internally collaborating by reusing the agents of other teams and sharing their own. A practical example we use internally involves a team of clinical researchers in R&D. They utilize a data retrieval agent from Development and a data visualization agent from Biostat, which communicate with each other through a common agentic framework. This collaboration helps them answer the question: “In all our previous trials, did semaglutide help patients reduce their smoking habits?” One agent fetches the data, the other visualizes it on the fly. A process that could take weeks or months, is now taking seconds. This is what unlocking and democratizing data at scale in an organization looks like.
Hyperight: Can you explain how agentic architectures differ from traditional AI approaches? Particularly in handling data from clinical trials and medicinal research.
Georgios Ilias: I would start by pointing out that generative AI is new. There is little at the moment one could reliably call “traditional” in the functional utilization of the technology, no manual of battle-tested methods or best practices. It is this void we’re trying to fill internally, and the agentic architecture and framework are prime examples of that work in practice. My presentation is titled “Dividing AI to Conquer Stability,” and that division is at the core of the solution. We need to control and optimize AI. We need to know its purpose and know what we are optimizing towards. If I asked you to “evaluate OpenAI’s GPT-4o,” your response would probably be “evaluate it on what task?” If I asked you to evaluate an AI agent that can only test Python code and respond with issues and suggested fixes, you’d have a clear expectation, goal, and target for your evaluation. Agents are AI, packaged in a way that enables quantification, iteration, and quality control. Agents with the ability to talk and coordinate with each other. Now picture a set of specialist agents performing their tasks. They communicate with each other in a common internal language, coordinated by an orchestrator. Task execution evolves from a fixed monolithic flow using generalist models. It transitions to a flexible set of quality-controlled steps performed by extensively tested specialists.
Bringing the focus back to healthcare and clinical trials, is a highly complicated and sensitive data corpus to utilize. Relying on generalist models and monolithic architectures is simply not conducive to the process, let alone scalable and robust. Generative AI is typically non-deterministic, inference can fail and an architecture is needed to gracefully account for that. Testing of AI must be considered, both for individual tasks and upon integration with the broader ecosystem. Continuous evaluation needs to be implemented, leading to iterative changes and stable growth. Transparency and enabling the user to steer AI during task execution are vital. And so is AI asking the user for feedback and direction when uncertainty is high. To formally consider an AI-centric application ready to undertake some of the high-stakes tasks in a field as sensitive as healthcare such steps are imperative. Fortunately, we recognized this need early at Novo Nordisk. This has led to our development of the agentic architecture and framework as a solution.
Hyperight: In your experience as an Engineer, what have been the biggest challenges in implementing AI solutions in healthcare environments?
Georgios Ilias: Healthcare is a challenging field in many ways. The considerable sensitivity of the subject matter means care is needed, security has to be high, compliance with best practices is critical and therefore the adoption of new technologies can be slow. Furthermore, prioritization of IT as a core internal competence in healthcare is a relatively new concept. While these are blockers to the rapid adoption of radically novel technologies like generative AI, Novo Nordisk has done well so far at adapting to the new status quo. This adaptability reflects our commitment to embracing innovation despite the challenges.
Another challenge is the presence of internal silos, which is more a function of the large size of companies in this industry rather than an issue inherent to it. That said, I have personally witnessed excellent progress on all fronts in recent years.
Hyperight: In your upcoming talk at the summit, you will discuss the need for robustness and scalability in AI applications. What strategies do you employ to enhance these qualities in agentic architectures?
Georgios Ilias: Simply put, the tenet of divide and conquer. Reduce complexity through the division of large, monolithic AI-centric processes into smaller, more agile steps or tasks. Instead of using generative AI directly, we incorporate it into agents. An agent has a unit of logic defined by a suitable genAI or ML model, accesses tools that enable its function, and can operate independently. AI model selection, tools, or structured outputs control the output. Agents contain tests for their own function and validation checks to quantify their performance both in terms of quality and speed. It has a task, clearly defined and described, and a predetermined interface for communication with other agents.
We coined the concept of “agent purity,” borrowing from the single-responsibility principle, the KISS principle, and the concept of pure functions in programming. It encourages the creation of agents that are as simple, self-contained, and task-specific as possible. Furthermore, we enable the execution of more complicated tasks without violating the purity principle by allowing agents to compose other agents. This composition follows a fractal-like structure. Finally, we have the concept of a Plan. In this approach, an orchestrator agent separates a user query into the steps required for its execution, based on the agents available to it. The orchestrator is adaptable. It can withstand and recover from failures in both the plan it generates and the steps it assigns to agents. It can try new approaches, and directly involve the user in the plan, as users in our system are also considered agents.
Given the above, any agentic solution that implements our framework is highly scalable. Agents can be continuously added to expand its capabilities without requiring modifications to the rest of the system. Moreover, different teams can implement agents independently and share them. These agents only need to obey the common interfacing requirements to communicate and coordinate with any other agent.
In addition to scalability, these systems are robust. They can withstand momentary failures in parts of the plan or execution and gracefully adapt. Furthermore, the addition of the user as a vital and integrated part of the process is crucial to stability, effectiveness, transparency, and user experience in general.
Hyperight: What role do you believe agentic architectures play in democratizing access to data insights in the healthcare sector?
Georgios Ilias: Agentic architectures and AI in general can help data democratization in multiple ways. First and foremost, they can help make data more FAIR, with emphasis on Findable and Accessible. Our company has collected incredible amounts of data over the years. But users can only find, quickly access, and understand that data in an asynchronous, scalable way if it is useful. Humans in the loop for such repetitive tasks are not only prone to mistakes but dramatically slow down the process. Consider the people who need insights from data the most: leaders, managers, stakeholders, and medical researchers. They are rarely the ones who know databases, code, or how to perform exploratory analysis on big data. Our data engineers have huge backlogs of requests for such insights. Through our internal partnership with FounData, we have shown how that process can be reduced to a single natural language query. This query is executed in seconds using agents that can access the data, visualize it, and even explain those visualizations, turning them into insights. Humans will remain in the loop, but our species needs to start focusing on what we excel at: quality over quantity. We should prioritize creative tasks rather than endless, repetitive backlogs.
Hyperight: Looking ahead, how do you envision the future of AI in healthcare, particularly with the advancements in agentic architectures?
Georgios Ilias: The future looks bright! I envision the optimization of processes that used to take months, into automated processes that take seconds. Data will be king. And through producing insights from data and automating/streamlining repetitive internal processes, we will free up our bandwidth and foster human creativity and experimentation. In the end, the hope is always that medicine will reach more patients, faster and at more affordable prices.
Tune into Georgios’ presentation on “Agentic Architectures: Divide AI to Conquer Stability,” at the NDSML Summit 2024. Don’t miss this opportunity to explore the future of agentic architectures with cutting-edge AI insights!
Add comment