Monitoring and responding to regulatory changes are essential for compliance in the pharmaceutical industry. Companies must keep track of regulations from global agencies like the FDA and the European Medicines Agency. This process often requires manually reviewing numerous regulatory websites and supporting documents in various formats, which can lead to compliance issues and increased costs. This blog discusses how Amazon Bedrock Agents and Amazon Nova Pro's Large Language Model (LLM) can simplify the process of regulatory intelligence in the pharmaceutical sector. By automating these tasks, the solution aims to improve efficiency and reduce the time needed for compliance, allowing companies to focus on their core activities.
Business challenge
Pharmaceutical companies are constantly tracking and analyzing the new changes introduced by global regulatory agencies such as the FDA, European Medicines Agency and others. This requires manually scanning hundreds of regulatory websites, supporting documents in PDF format and then summarizing and documenting the content to assess the impact on their portfolio of drugs. Most organizations implement these steps manually, leading to compliance gaps, escalated cost and effort burden.
Key solution capabilities
To address these challenges, HCLTech developed a cutting-edge Agentic AI regulatory intelligence solution to streamline the entire regulatory monitoring and analysis process. The solution's key business characteristics are listed below.
- Automated web crawling across multiple regulatory agency websites and their sub-sites
- Smart content summarization and report generation from complex regulatory documents
- Secure data storage with interactive chat-based query capability
- Seamless integration with existing Regulatory Intelligence Management (RIM) systems
- Complete traceability with source documentation and audit trails
Solution architecture on AWS
Amazon Bedrock: A fully managed service that provides access to Foundation Models (FM) through a simple API interface.
Amazon Bedrock Agents: The solution utilizes the Amazon Bedrock Multi-Agent framework through a coordinated system of four agents: one orchestrator agent (Supervisor agent) and three collaborator agents: a Crawler agent, a Summarization agent, a Notification agent and an Integration agent.
Large language model: The agents leverage Amazon Nova Pro as a foundation model.
Amazon Bedrock knowledge bases: To provide contextual information for a company, from its private data sources.
Amazon API gateway WebSocket APIs: This acts as a persistent chat interface to route communication between users and Bedrock Agents via AWS Lambda.
AWS Lambda: User input is received via the WebSocket API and sent to Bedrock Agents by calling the invoke Agent function. AWS Lambda is also used as part of the Bedrock Agents action group.
Other AWS services used include SNS and S3.
Solution architecture: Request flow in a multi-agent architecture pattern
- The end user provides input on regulatory sites that must be crawled to gather information. This can be done either by passing the website data via UI or as a scheduled event, which can be triggered using AWS EventBridge.
- These details will be sent to an AWS Lambda function via the AWS WebSocket API. The WebSocket API maintains a persistent connection between the user and Bedrock Agent for querying.
- The AWS Lambda function invokes the boto3 Agent API and passes the inputs to the Amazon Bedrock Agent Service. The Bedrock Agent Service has four agents configured: an Orchestrator agent, a Crawler agent, a Summarization agent, a Notification agent, and an Integration agent. The Orchestrator agent is used for routing and collaboration between various agents.
- Once the Orchestrator agent receives a user request for particular websites to be crawled via the invoke Agent API, it routes the request to the Crawler agent. The Crawler agent invokes the Action Group with AWS Lambda, and the sites are crawled using the Tavily search. Once the data is crawled, it is sent back to the Orchestrator agent.
- The orchestrator agent will then route it to the summarizer agent to summarize the content. After generating the final report, the Summarizer agent will send it back to the Orchestrator.
- This is then further routed to the Notification and Integration agent. This agent can load the final report to an S3 bucket and notify the user using AWS SNS.
- Another requirement from Pharmaceutical companies is how the output generated can be integrated with the existing RIMS solution. This can be done with the Amazon App Flow service, which allows easy integration with SaaS-based services.
- Finally, if Q&A crawled data is required for further insights, it can be written to an S3 bucket and vectorized using Bedrock Knowledge Bases, making it easy to access via the chat interface.
Security architecture and Responsible AI
The solution has a robust security architecture and incorporates Responsible AI principles.
| Security domain | Implementation strategy and AWS services |
|---|---|
| Responsible AI – Toxicity, PII redaction, hallucination | AWS Bedrock guardrails for toxicity control, PII redaction and contextual grounding to detect and mitigate hallucinations |
| Data at rest | AWS Key Management Service (KMS) for encrypting data in Amazon S3. |
| Data in transit | AWS PrivateLink with VPC interface endpoints for secure AWS service-to-service communication; TLS encryption for all client-service interactions |
| Authentication and authorization | Amazon Cognito (or similar enterprise solution) for secure user identity management and authentication flows; IAM for fine-grained permission control with role-based access |
| DDoS protection | Multi-layered defense with AWS Shield for network layer (L3/L4) protection and AWS WAF for application layer (L7) filtering against sophisticated attack vectors |
Benefits of using the Agentic AI solution
The solution can help accelerate the pace at which Pharmaceutical companies can comply with the changing regulations with considerably lower effort. The solution can easily scale and allow Pharmaceutical companies much deeper insights through a chat interface.
Additional points to consider during implementation
- Certain websites have a large volume of data when crawled using Bedrock Agent action group—AWS Lambda can hit the maximum size limit of 25KB while returning crawled data. For those scenarios, once the data is crawled using AWS Lambda, the response can be sent to Nova Pro LLM to summarize and then sent back to the Agent.
- Certain payload size limitations apply to the AWS WebSocket API on input or output data that can be sent. This is mainly applied for security reasons to block malicious content that can overwhelm the application with a large payload size. To overcome this, the input or output file generated can be uploaded to an S3 bucket, and a Pre-signed URL can be sent to the user via a chat interface to access the data.
Conclusion
The solution demonstrates building a comprehensive regulatory intelligence platform using Amazon Bedrock Agents and blazing-fast Nova-Pro LLM from AWS.
For more detailed information, a demonstration, or to implement this solution, please reach out to our team of experts at awsecosystembu@hcltech.com
