In Part 1, we covered how we defined the consumer problem for Nayya Claims Advocate, how we rapidly prototyped an MVP, and concluded with how we took the prototype to a full-fledged product.
Building agentic AI products presents a series of challenges. Hallucinations refer to seemingly plausible, yet entirely fabricated or erroneous, presentations of information from an LLM. In the Nayya Claims Advocate, we took a number of steps to mitigate the potential for hallucinations, these include: advanced prompting techniques such as reflection; a heavily curated knowledge base for document retrieval; and a “QA agent” whose job it is to ensure that the content of appeal letter matches all of the contextual information.
Data privacy and security are paramount for all applications, particularly those in the healthcare space. When working with large language models (LLMs), these concerns become even more complex. Unlike traditional software components, LLMs can inadvertently memorize and surface sensitive data observed during training or inference. Their black-box nature makes it more challenging to audit and enforce strict data handling rules. Moreover, prompts and responses can contain protected health information (PHI), making it essential to tightly control where and how that data is processed and stored.
To address these challenges, we chose to use AWS Bedrock, which allows us to leverage foundation models in a secure, managed environment without exposing sensitive data to external providers or open APIs. Bedrock runs within our existing AWS infrastructure, enabling us to enforce encryption at rest and in transit, restrict access via IAM roles, and log all model interactions centrally. Paired with strong software engineering compliance practices—such as rigorous input/output validation, role-based access controls, data minimization, and automated auditing—we're able to meet strict compliance standards. This approach helps ensure that, while we build cutting-edge AI applications, we do so in a way that prioritizes security and user trust.
Large Language Models (LLMs) are black boxes, meaning that the internal workings of the models are, in practice, impossible to communicate to the user in real-time. Since the Nayya Claims Advocate deals with highly sensitive information and has high-stakes consequences, we have built in several ways for the user to understand the agent’s thinking and provide any necessary corrections along the way. Furthermore, we developed ways for the agent to explain how it produced certain portions of the letter, for example, the research it conducted to find relevant clinical information. These methods of explainability improve trust while allowing users to provide real-time feedback if the agent is heading in the wrong direction.
Building the Nayya Claims Advocate, Nayya’s first fully agentic AI product, was an incredible journey—from concept to beta in 12 weeks. It showed us that the future of AI products isn’t just in models—it’s in building innovative, secure, well-integrated systems around them. As we look ahead, we’re excited to apply what we’ve learned to a growing set of use cases where autonomy, explainability, and trust are just as important as raw capability. Most importantly, this product isn’t just a technical accomplishment—it’s a step toward making the healthcare system more accessible and humane, which will lead to human thriving, our ultimate mission. That’s the kind of impact we want every AI product we build to have.