AI Architecture
Our approach to AI focuses on using smaller, yet highly capable models to tackle big challenges like compliance and security, with a per-customer architecture. ThetaRho’s AI platform goes beyond just models and infrastructure—it offers a unified stack for experimentation, development, testing, deployment, and support. This enables faster iteration of end-to-end solutions.
Each customer receives their own dedicated GenAI model, ensuring models aren’t shared across customers. Additionally, no personally identifiable information (PII) is ever sent to public GenAI services like ChatGPT, maintaining strict data security and privacy. We leverage an ensemble of models that work together to classify, route, embed, generate natural language responses, and validate the output of GenAI models, all integrated into what the industry calls an “Agentic Workflow.”
Data Architecture
Data is the foundation on which all AI solutions are built, Healthcare is no exception. However, there are many challenges that are specific to the healthcare industry. Some of these are:
Healthcare’s data problem
The Healthcare industry not only produces 30% of all the data generated across all industries, this data is also growing faster than other data-driven industries like finance and media. This is not that surprising with the explosion of digital health devices. However, as an industry, it tends to be a late adopter when it comes to adopting new technologies because of the perceived complexity of complying with regulatory requirements in the industry.
What is worse is 97% of data generated by hospitals goes unused. This data includes clinical notes, lab tests, medical images, sensor readings, genomics, operational and financial data.
Current EHR systems, which serve as the primary system of record in hospitals, are primarily optimized for billing rather than patient care. This often results in fragmented data from a user perspective, even within a single organization. The challenge lies in the fact that different data technologies are required for different use cases, and once a foundational data technology is chosen, it becomes difficult for organizations to adopt newer technologies without major operational disruptions. To address this, ThetaRho has developed its own data layer that integrates with, but remains separate from, EHR data, offering greater flexibility and efficiency.
Semantic Store as key enabler
However, with the widespread adoption of FHIR and the availability of GenAI, there is a real chance for providing a natural language interface for users to access all the data that is relevant to patients when needed.
For this to happen, the same data potentially needs to be available in multiple representations. For example, if we need the result of a natural language query to provide an integrated timeline that combines test results, medications and conditions in a single report, an analytic store is needed. If we want to provide ability to query the patient record using natural language, we may need a search index as well as a vector store for storing embeddings of fragments of data from EHR. Alternatively, if we want to ground output of LLMs, we may need a graph store for storing domain dependent knowledge graphs. Not all these representations are needed for every use case, but each of these solves the data problem for a particular use case. We call this collection of data technologies our semantic store because it enables us to present the appropriate information to the user in the right form factor depending on the use case.
