SignalCare

Natural-language symptom descriptions are useful but inconsistent. A healthcare-oriented prototype also cannot let a generative model silently control high-stakes escalation.

The system needed to demonstrate useful extraction and follow-up behavior without presenting itself as a diagnostic product or delegating emergency decisions to model output.

• Model output is always treated as untrusted input.
• Emergency red flags must remain deterministic and inspectable.
• Region-aware terminology must not be confused with medical validation.

The React intake surface sends narrative input through a serverless boundary. Structured extraction and adaptive follow-ups sit beside a deterministic rules layer that owns emergency escalation and the explainable care-level result.

The central decision was separating language understanding from authority. The model can extract, summarize, and draft; deterministic rules can override and escalate.

The flow covers natural-language intake, structured symptom extraction, adaptive follow-ups, region-aware terminology, an explainable care level, a provider-ready summary, exports, and an evaluation dashboard.

Red-flag decisions are outside the generative model. The interface exposes the deterministic override path and keeps the prototype's limitations visible.

The included deterministic suite contains 33 cases. All 33 pass, with zero emergency misses in that local suite. These project-local results are not third-party or clinical validation.

The prototype demonstrates how an AI-assisted interface can remain useful while deterministic verification retains control of high-risk outcomes.

SignalCare is educational software. It is not a medical device, diagnosis tool, prescription system, emergency-care replacement, or clinically validated service.

Reliability often comes from narrowing what a model is allowed to decide, not from writing a more persuasive prompt.