Scaling Generative AI in pharma requires a flexible, high-performance architecture to move from proof of concept to enterprise-wide deployment. (Image: Adobe Stock)
Scalability is a common issue facing pharma leaders who have deployed Generative AI (GenAI) applications as a proof of concept (PoC). Many are dazzled by their promise of streamlining staid work processes. Agentic AI and other new GenAI technologies can make PoCs faster and easier to launch, advancing scalability as well as security and development speed.
Yet, when the organization moves beyond the PoC to broad-scale deployment, applications often cannot perform with larger numbers of users or debut with little to no adoption. Small but crucial elements of the workflow create barriers to completion.
Achieving the innovation that the PoC promises requires more than the latest GenAI technologies. It requires a common-sense strategy that builds longstanding organizational-process expertise into the workflow. In this manner, teams build flexibility into GenAI applications so they can scale and deliver the improvement gains promised, and sometimes even more.
Integrating subject-matter experts (SMEs) for accuracy
Once teams establish that the PoC delivers business value, defining the goals and success metrics for the rollout begins. Integrating subject matter experts (SMEs) into the development process at this stage is critical to training and testing the AI output.
It’s not enough to have a highly skilled IT and R&D team available to explore advancements and prioritize AI implementations. Pharma organizations need to create internal AI knowledge repositories where teams share best practices and lessons learned from their AI projects. Establishing regular AI knowledge-sharing sessions with cross-functional resources helps to constantly share ideas across the organization.
It is also important to incorporate human-in-the-loop (HITL) validation that mitigates potential rogue conclusions. For example, in systematic literature reviews (SLRs), human-in-the-loop workflow, assessment, and validation are essential to managing decision points. Best practice in the industry is for AI-aided literature review solutions to provide transparency and explainability by offering additional steps that help users understand and evaluate the reliability of predictions made by the AI models, along with visibility into when AI is used in the literature review.
Flexible, high-performance scalable architecture
Another pivotal requirement is to design the system using new scalable AI components and a microservices architecture to help ensure performance.
Pharma teams can create a secure and scalable API infrastructure to serve as a foundation for scaled-out deployment. Modular design and flexible architecture are important since they allow for easy updates and modifications without requiring complete system overhauls. The ability to swap in different large language models (LLMs)—for example, Gemini, ChatGPT, Claude, among other models—while maintaining a stable core structure is essential for systems to maintain flexibility.
A common approach is to combine cloud native services, autoscaling, serverless architecture, caching, container orchestration, RAG, CI/CD pipelines, MLOps, and model monitoring among other techniques. This setup ensures optimal performance, cost efficiency, and high availability for cloud-native applications.
A Web Application Firewall (WAF) acts as the first layer of defense, protecting APIs from common web threats such as SQL injection, cross-site scripting (XSS), and distributed denial-of-service (DDoS) attacks. The WAF inspects incoming traffic, applying rules to block malicious requests before they reach the API Gateway. This ensures only legitimate requests are processed, enhancing overall security.
Often an API Gateway serves as the entry point for all API requests to enable routing, throttling, and authentication while providing a managed solution to handle millions of requests per second. With the API Gateway, applications can enforce:
- Rate limiting to prevent abuse.
- Caching to reduce repeated processing of similar requests.
- Request validation to ensure input integrity before forwarding to backend services.
Most flexible architectures for AI development also use Retrieval-Augmented Generation (RAG), which allows AI to pull real-time, relevant external data while ensuring data gathered remains up to date with the latest information.
Kubernetes provides a robust solution for horizontal scaling by automatically adjusting microservice instances, while AWS Lambda allows for efficient handling of event-driven tasks like image processing, ensuring both cost-efficiency and high availability.
To handle surges in traffic, GenAI application developers also usually add:
- Load Balancing, to distribute traffic evenly across multiple instances.
- Read Replicas, to reduce the load on the primary database.
- Rate Limiting and Request Queuing, to prevent system overload.
In addition, continuous monitoring and autoscaling can help adapt to workload variations, keeping the system resilient and performant under any condition.
Allowing LLM flexibility and validation
Using Agentic AI to do the initial validation with different personas can help minimize the human reviews. Agentic AI has become an AI buzzword and rightly so, as it helps automate specific granular workflows and combine them as required to keep decision-making adaptable to new insights and shifts in policy.
In drug development, modular AI systems enable life sciences companies to use the best-suited LLMs for each task, for example Claude might be chosen for data extraction and ChatGPT-4.0 for clinical synthesis, allowing teams to adapt quickly to innovation. This flexible approach supports faster drug development, seamless integration of new tools, and optimized collaboration across specialized AI components like data ingestion, logic, and user interfaces.
Security and privacy
Last but certainly not least is security. AI infrastructure should follow zero trust principles, which means that every AI request or action is verified before execution. At the hardware level, System on Chip (SoC) security frameworks also can provide AI-powered security. Best practice is to encrypt data at rest and in transit, incorporate identity and access management (IAM) such as two-factor authentication, and use audit logs, as well employing additional security measures like Adversarial Attack Resistance, Secure Model Storage and Versioning, Data Governance for AI, and Configuration Management.
Taking efficiency beyond the PoC
Gaugarin (“G”) Oliver
AI has evolved rapidly in the past few years, and with underlying AI technologies updated at least monthly, organizations must expect continual change. When structured and deployed with forethought and flexibility, the AI deployment can continue to scale and deliver the innovative time- and cost-savings that the PoC demonstrated, and sometimes even more.
About the author
Gaugarin (“G”) Oliver is the founder and CEO of CapeStart, Inc., a professional services firm supporting the healthcare and life sciences industry, and maker of the award-winning MadeAi platform that helps pharma organizations compete and win in the AI economy.
Filed Under: Drug Discovery and Development, machine learning and AI
