{"resourceId":"unc-sarhachat-openshift-ai-case-study-20260916","versions":[{"version":"external-2b5695e6d7c6d3ec839d6cac5bd1318f71861d3cfc0cc0cb97f6aebd3d038535","resource":{"id":"unc-sarhachat-openshift-ai-case-study-20260916","title":"University of North Carolina builds human-centered AI assistant (SARHAchat) on Red Hat OpenShift AI","organization":"University of North Carolina at Chapel Hill (School of Nursing, Computer Science, RENCI)","sector":"Higher education / public health information","geography":"Chapel Hill, North Carolina, U.S.","publishedAt":"Red Hat customer case study, Sept. 16, 2026","publicationDate":"2026-09-16","eventDate":null,"sourceName":"Red Hat","sourceLabel":"University of North Carolina OpenShift AI case study","sourceUrl":"https://www.redhat.com/en/resources/university-of-north-carolina-openshift-ai-case-study","evidenceClass":"vendor-claim","outcomeClass":"emerging","topics":["knowledge-work","infrastructure","accessibility-workforce"],"finding":"Vendor claim: UNC's School of Nursing, Computer Science department and RENCI moved SARHAchat, a sexual and reproductive health assistant for 18 to 25-year-olds, from a general-purpose LLM prototype built over 2 years to Red Hat OpenShift AI. Working with Red Hat pre-sales, the team built a prototype in 2 weeks. It reports smooth switching between informational and consultative modes, safety checks for basic health screening, and a move from a model of more than 500 billion parameters to a 20-billion-parameter model, which the team says gives faster responses on less infrastructure. The system and data are hosted on premises on NVIDIA hardware. Public testing and interviews are still under way, and the long-term aim is handoff to non-profit organizations.","sledRelevance":"A public university case of on-premises generative AI for sensitive health information. It is relevant to university research IT, academic medical centers and state public-health agencies weighing self-hosted, smaller models for sensitive-topic assistants. The planned handoff to non-profits highlights sustainability and maintainability questions that SLED programs face after grant-funded pilots.","evidence":"Read the web case study (datePublished 2026-09-16) and the linked PDF. All metrics come from vendor marketing with customer quotes: a prototype in 2 weeks and a 500B-to-20B model reduction. No user outcomes, accuracy, safety evaluation results, cost figures or usage numbers are reported. The assistant is still being tested with members of the public, so this is a pre-production result.","architectureImplications":"On-premises OpenShift AI on NVIDIA hardware with model swapping (the team is exploring Google Gemma 4). Right-sizing to a 20B model reduces GPU requirements for serving.","governanceImplications":"On-premises hosting is cited as giving control that helps align with CDC guidelines. Clinical safety governance for a health assistant is described only qualitatively.","securityPrivacyImplications":"Sensitive reproductive-health conversations stay on university infrastructure, but the case study gives no detail on logging, retention or access controls.","caveats":"Vendor-published case study, pre-production, with no independent evaluation or quantitative outcomes beyond model size and prototype speed. No event date is given for the migration.","streamIds":["red-hat"],"roles":{"sales":"Sales framing: a credible public university story for sensitive-domain assistants on self-hosted OpenShift AI. The best hook is the shift from a giant general-purpose model to a right-sized 20B model on premises. That speaks to cost, data control and maintainability for universities, academic medical centers and public-health agencies. Be careful with claims. This is pre-production, vendor-published, and reports no user or safety outcomes. Position a scoped prototype with an evaluation plan, not proven clinical value.","engineering":"Engineering framing: the reusable pattern is evaluating and right-sizing models on OpenShift AI, then serving a smaller model on premises with a dual-mode conversation design and safety screening built into the application. For similar work, design model-swap capability from the start, add evaluation harnesses for accuracy and safety before public testing, and size GPU capacity for the target model, not the prototype's. Logging and retention for sensitive conversations need explicit design.","delivery":"Delivery framing: the case shows a fast prototype, but production readiness depends on work still in progress, including public testing, interviews and iteration. Plan clinical and ethics review, accessibility testing for the planned voice and video features, and a documented operating model if the service will transfer to a non-profit or other operator. Define support ownership, model update procedures and success measures before scaling beyond the pilot."},"roleDetails":{"sales":{"customerProblem":"Institutions want trustworthy, private AI assistants for sensitive topics without depending on large third-party models whose behavior changes with upgrades.","stakeholders":["Academic or clinical program lead","Research computing (e.g., RENCI-like groups)","University CIO/CISO","IRB or clinical safety reviewers"],"discoveryQuestions":["What sensitive domain would the assistant serve, and who reviews its safety?","Do you have on-premises GPU capacity?","Who will operate the service after the pilot or grant ends?","How will you measure accuracy and user benefit?"],"valueHypothesis":"Right-sized, self-hosted models on OpenShift AI may lower serving infrastructure needs and provide more stable behavior than general-purpose hosted models.","potentialEngagement":"Model right-sizing and evaluation pilot on OpenShift AI for one sensitive-domain assistant.","unsupportedClaims":["Any health outcome or user benefit","Cost savings amounts","That the assistant is safe for clinical use"]},"engineering":{"fit":"Fits institutions with on-premises GPUs and OpenShift that need data control for sensitive conversational AI.","architectureIntegration":"OpenShift AI model evaluation, fine-tuning and serving on NVIDIA hardware; application-level safety screening and dual-mode conversation design.","prerequisites":["On-premises GPU capacity","OpenShift AI subscription","Domain-expert evaluation datasets"],"constraints":["No published accuracy or safety evaluation","Pre-production status","Vendor-only metrics"],"security":"Keep conversation data on premises with defined logging, retention and access controls; the case study does not describe these.","proofOfValue":["Compare candidate models on a domain evaluation set before selection","Measure response latency and GPU use for the chosen model","Run a safety review of screening logic with clinicians"]},"delivery":{"work":"Prototype, evaluate and harden a sensitive-domain assistant, then define its operating and handoff model.","dependencies":["Clinical and ethics review","GPU availability","Operator for long-term support"],"ownership":"Academic program owns content and safety; research computing owns platform operations.","skillsAdoption":"Train program staff on model evaluation and the operating team on OpenShift AI serving.","governanceCheckpoints":["Safety and ethics review before public testing","Accessibility review of new interfaces","Handoff readiness review"],"acceptanceCriteria":["Proposed: documented accuracy and safety evaluation results","Proposed: defined retention policy for conversations","Proposed: named operator and support model before scale-up"],"risks":["Unsafe or inaccurate health guidance","Sustainability gap after pilot funding","Privacy exposure from conversation logs"]}},"retrievedAt":"2026-10-30T03:09:23Z","enrichedAt":"2026-10-30T03:21:15Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"The team plans voice and video features to reach a wider audience and stresses that operating the system must be manageable for organizations without in-house AI expertise.","procurementImplications":"The work was done with vendor pre-sales help. Institutions should budget for post-pilot support and subscriptions if a non-profit or other operator takes it over.","operatingModelImplications":"Shows a university-built service intended for transfer to external operators, which needs a clear support and maintenance model.","sourceVerification":{"openedUrl":"https://www.redhat.com/en/resources/university-of-north-carolina-openshift-ai-case-study","referenceExcerpt":"The team reduced the model size from more than 500 billion to 20 billion parameters, increasing efficiency and response times.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}