The challenge
Athirion AI is a healthcare technology company in East Brunswick, New Jersey, running an AI powered clinical operations platform for healthcare providers. It spans four areas: clinical workflow automation, patient intake, HIPAA compliance monitoring, and revenue cycle intelligence, built on AWS with an HL7 FHIR native data layer and integrations across EHR, practice management and payer systems.
The team had already selected Claude on Amazon Bedrock and was preparing to move its first Claude backed features toward production across prior authorization, patient intake, compliance monitoring and denial management. The architecture and the model access decision were settled.
What did not exist yet was the working practice to build on them. There was no shared structure for prompt design, no test set or scoring method to tell whether a prompt change improved or degraded output, and no written internal standard for what may safely enter a prompt or a log in an environment handling protected health information. In a regulated clinical setting, establishing that discipline before the first release is materially cheaper than retrofitting it afterward, which is why the program ran when it did.
What Kriv AI did
A curriculum built on Athirion AI's own workflows
Three weeks of instructor led sessions delivered remotely to a cohort of up to ten people across engineering, product and operations. Every teaching example used Athirion AI's own terminology, payer rule structures and output schemas rather than generic material, and three priority workflows chosen during week one discovery carried through the whole program.
- Roughly ten hours of live delivery across five sessions: Claude fundamentals for clinical operations, prompt engineering for clinical accuracy, structured outputs and tool use, evaluation and quality control, and safe deployment practices.
- Three hands on labs run by the participants themselves inside Athirion AI's own environment, against the three priority workflows.
- Office hours alongside the sessions for troubleshooting, prompt review and applied questions.
A prompt library and an evaluation kit they keep
The durable half of the engagement. The point was not that the team had working prompts at the end of three weeks, it was that they had a repeatable way to tell whether the next prompt was better or worse than the last one.
- A prompt library covering the three priority workflows, each prompt documented with intent, input assumptions, output schema and known limitations, executable against their own Bedrock environment.
- An evaluation starter kit: test case template, scoring rubric, defined pass and fail thresholds, and a worked example run end to end. That gives the team a release gate for Claude backed features rather than a judgement call.
- Structured output patterns producing schema conformant JSON for downstream FHIR and claims processing.
- A safe use standards summary and a prioritized 90 day adoption roadmap.
De-identified in, no production access
The security model was set before delivery, not negotiated during it. Kriv AI received no protected health information at any point and held no credentials to Athirion AI production systems or data stores. Every lab ran inside Athirion AI's own AWS tenant against de-identified or synthetic data the customer supplied. The safe use standards defined minimization of protected health information in prompts and logs, handling of untrusted input, and mandatory human review points in Claude assisted clinical workflows.
Results
- Staff across engineering, product and operations moved from no structured practice for Claude development to building and evaluating Claude workflows independently.
- Claude is in day to day use at Athirion AI across clinical operations work.
- The prompt library, evaluation kit, safe use standards and 90 day roadmap stay in house as a durable capability base.
- The team was operating independently at handover, with no ongoing dependency on Kriv AI.
About Athirion AI
Athirion AI is a healthcare technology company headquartered in East Brunswick, New Jersey, operating an AI powered clinical operations platform for healthcare providers. The platform covers clinical workflow automation, patient intake, compliance monitoring and revenue cycle intelligence, built on AWS with an HL7 FHIR native data layer. For more information, visit athirion.ai.