Bayer AG, in collaboration with Thoughtworks, developed the Preclinical Information Center (PRINCE) to modernize data retrieval and analysis in pharmaceutical research. The platform transitioned from a keyword-based search tool for structured metadata into an agentic AI system capable of handling complex queries across decades of unstructured PDF study reports. This evolution occurred in three phases: Search, which consolidated fragmented data silos; Ask, which implemented Retrieval-Augmented Generation (RAG) to allow natural language queries; and Do, which utilizes multi-agent workflows to automate research tasks and draft regulatory documents.

PRINCE operates on a backend orchestrated by LangGraph, with a user interface built in React. The architecture employs specialized agents—a Researcher, a Writer, and a Reflection Agent—to manage distinct stages of the request process. These agents retrieve information from OpenSearch for unstructured vector data and Athena for structured datasets. To ensure reliability, the system tracks state using a LangGraph checkpointer persisted in PostgreSQL and DynamoDB. The platform integrates models from OpenAI, Anthropic, Google, and open-source providers via a unified endpoint, applying context discipline to ensure each agent receives only the information necessary for its specific function.

Resilience is maintained through automated fallbacks that retry individual LLM calls or entire logical nodes if failures occur. The system incorporates observability via Langfuse and Cloudwatch to monitor performance and debug issues using the RAGAS evaluation framework. By implementing proactive clarification of user intent and human-in-the-loop validation, PRINCE aims to reduce manual analysis time and improve research efficiency. The platform design prioritizes transparency and governance, allowing researchers to accept or override AI-suggested data sources, thereby maintaining human oversight in preclinical development workflows.

Bayer AG, in collaboration with Thoughtworks, developed the Preclinical Information Center (PRINCE) to modernize data retrieval and analysis in pharmaceutical research. The platform transitioned from a keyword-based search tool for structured metadata into an agentic AI system capable of handling complex queries across decades of unstructured PDF study reports. This evolution occurred in three phases: Search, which consolidated fragmented data silos; Ask, which implemented Retrieval-Augmented Generation (RAG) to allow natural language queries; and Do, which utilizes multi-agent workflows to automate research tasks and draft regulatory documents.

PRINCE operates on a backend orchestrated by LangGraph, with a user interface built in React. The architecture employs specialized agents—a Researcher, a Writer, and a Reflection Agent—to manage distinct stages of the request process. These agents retrieve information from OpenSearch for unstructured vector data and Athena for structured datasets. To ensure reliability, the system tracks state using a LangGraph checkpointer persisted in PostgreSQL and DynamoDB. The platform integrates models from OpenAI, Anthropic, Google, and open-source providers via a unified endpoint, applying context discipline to ensure each agent receives only the information necessary for its specific function.

Resilience is maintained through automated fallbacks that retry individual LLM calls or entire logical nodes if failures occur. The system incorporates observability via Langfuse and Cloudwatch to monitor performance and debug issues using the RAGAS evaluation framework. By implementing proactive clarification of user intent and human-in-the-loop validation, PRINCE aims to reduce manual analysis time and improve research efficiency. The platform design prioritizes transparency and governance, allowing researchers to accept or override AI-suggested data sources, thereby maintaining human oversight in preclinical development workflows.