Published Paper Details:

Placing a large language model in front of real patient data in a multi-tenant clinical system creates three problems at the same time. The model can invent clinical numbers, it can expose protected health information, and the volume of text sent to it tends to grow without any natural bound. This paper describes the Fact-Curated Helper (FCH), an architectural pattern we built into Rudhiraksh, a multi-tenant thalassemia and blood-bank management platform that is in live use at a partner blood bank. The central idea is to keep computation out of the model. A deterministic layer, scoped to the caller's role and tenant, computes every clinical fact in SQL and assembles a small JSON object containing only what the caller is permitted to see; the model is then asked to phrase that object in plain language and to do nothing else. We describe three parts of the design. The first is a separation of computation from narration that prevents the model from producing clinical values. The second is a role-aware and tenant-aware grounding step that enforces access control in code before the model is called, so that no prompt can widen the model's view. The third is a three-tier provider stack (Gemini, then Groq, then a deterministic rule engine) that keeps the feature available even when a provider is down. On the deployed system the context stays compact and grows with the caller's authorisation scope rather than with the size of the database: a patient context is roughly 260 to 300 tokens and an administrator's context grows by about 57 tokens for each additional blood bank. In a controlled experiment that calls a hosted model under three prompting conditions, FCH reduces the per-patient prompt by about 55 percent against a naive full- record baseline (measured with the provider's own tokenizer) and removes 23 of 24 identifying fields from anything the model receives, with no loss of accuracy on the questions tested.