Cutting results-posting time 94% using AWS with Xogene

Registering a clinical trial protocol, posting results, and preparing plain-language summaries require experts to find relevant information in complex source documents and present it in the format that each registry requires. Before Transcend, specialists reviewed thousands of PDF pages and extracted data manually. Drafts then passed through quality control, quality assurance, study team review, and sponsor approval. A single results posting could take close to 69 hours of work on average, including responses to registry comments.

Plain-language summary requirements added another layer, with expert writers translating complex clinical trial results into language that patients, caregivers, and other readers could understand. Reducing that workload required more than just faster processing. Xogene needed a way to minimize repetitive drafting and checking while keeping specialists in charge of source selection and output quality.

Xogene was already using AWS for scalability, security, and access to multiple AWS Regions, so Amazon Bedrock was a natural choice for building generative AI applications and agents at production scale. The service gave Transcend’s developers the flexibility to switch between multiple large language models in a single environment as needed. In addition, using the security features of Amazon Bedrock, the company can protect sensitive data under the AWS Shared Responsibility Model.

The team also uses the Amazon Bedrock Converse API, which provides a consistent interface that works with all models that support messages. This common interface for calling models from different providers, including Anthropic’s Claude, simplified integration as developers tested different models.

Using LangChain and LangGraph, the Xogene team built Transcend as a multiagent system, including a supervisor agent that routes tasks to specialized agents for drafting and quality review. The solution uploads documents to Amazon Simple Storage Service (Amazon S3), which is built to retrieve virtually any amount of data from anywhere. For plain-language summaries, Transcend provides AI-generated explanations and source references so that reviewers can check drafts against the source material.

A single agent couldn’t ingest source material and produce output at the required quality. The team divided the work into tasks that mirror how human specialists work—drafting, fact-checking against source documents, and formatting for jurisdiction-specific requirements. “We’re treating AI like we would a junior writer,” says Woo Song, cofounder and CEO at Xogene. “So again, the final judgment belongs to the human expert.”

In a results-posting use case, Xogene cut total processing time from about 69 hours of manual work to around 4 hours, a 94 percent reduction. Drafting time fell from about 40 hours to around 1 hour, a 98 percent decrease, and quality control time dropped from about 10 hours to 30 minutes, a 95 percent reduction.

In a results-posting proof of concept (POC) with a top-five pharmaceutical company, Transcend delivered 95.2 percent average field-level accuracy, 100 percent completeness, and no observed hallucinations in the evaluated output. Xogene reported that the authoring portion took under 1 hour, compared with 70 hours under the company’s own authoring process. In its second run, a separate protocol-registration POC with the same company recorded 97.7 percent average field-level accuracy, with no observed hallucinations, and saved 15 hours of manual work.

Xogene has also expanded Transcend’s use to a client’s public AI assistant, which answers patients’ questions about clinical trials and helps them determine their eligibility. “With support from the AWS team, we could work through technical issues as we developed our agentic AI workflows,” says Song.