Why NOX Built Leo
Nox built Leo because a health conversation can include signs of an emergency, and an answer should never take priority over urgent safety guidance. Leo is Nox’s user-visible medical-safety system: it screens messages for acute clinical red flags before the AI model responds and surfaces guidance to seek appropriate care when needed.
Why does health AI need a safety guardian?
People do not always arrive with a neatly labeled health question. They may describe symptoms in everyday language, mention a change that feels alarming, or ask whether they should wait and see. In some cases, those descriptions can contain signals of an acute emergency.
A general health answer is not the right first response when a message suggests a potentially time-sensitive situation. The safer design is to recognize the risk early, put urgent guidance first, and avoid letting a conversational response bury what matters most.
That is the purpose of Leo. It is not a diagnosis tool and it does not replace a clinician or emergency services. It is a safety net designed to screen conversations for medical red flags before Nox provides AI-generated content.
What problem was Leo built to solve?
Health AI can be helpful for plain-language questions about symptoms, wellness, nutrition, medications, sleep, and everyday habits. But helpfulness is not the same thing as safety. A clear, plausible answer can still be inappropriate if someone needs urgent evaluation instead.
Leo was built around a simple priority: when a message may indicate an emergency, the system should surface seek-care guidance before the conversation continues. That approach recognizes that medical safety is not a disclaimer placed at the bottom of an answer. It needs to be part of the system’s behavior before an answer is generated.
This distinction also matters because people do not always use clinical terms. They may use slang, indirect phrasing, an older disease name, or a language other than English. A safety system has to account for how people actually describe urgent experiences, not only for a short list of exact phrases.
For a closer look at this principle, read Why Health AI Should Escalate Instead of Diagnose.
How Leo screens health conversations before an answer
Leo uses a layered approach. Its first check is an independent, deterministic detection layer that runs before any AI model is called. This layer uses fixed rules to screen every message for acute red flags.
The deterministic triage layer includes more than 100 rules across more than 70 acute red-flag categories. These include signs associated with stroke, chest pain and cardiac symptoms, severe breathing difficulty, severe allergic reactions, trauma and severe bleeding, mental-health crisis, pregnancy warning signs, concerning symptoms in children, environmental emergencies, and toxicology or overdose.
When one of those rules fires, Nox displays a clear emergency or urgent-care banner before any AI-generated content. The sequencing is deliberate: safety guidance is not appended as an afterthought or left to the model to decide whether to mention.
When the deterministic layer does not find a match, a second check can provide an additional backstop. A lightweight AI classifier re-reads recent messages to look for dangerous descriptions that fixed patterns may miss, including indirect language, slang, another language, or older terminology.
That backstop can add a safety note when it identifies a likely emergency. It cannot remove, weaken, or override a warning from the deterministic layer. Learn more about the design in How Leo Works.
Why use deterministic rules alongside AI?
AI can interpret flexible language, but its responses are not deterministic: the same kind of prompt is not handled by a fixed, fully predictable process. For high-stakes red-flag screening, Nox designed Leo so an independent rule-based check happens first.
A deterministic layer makes the initial screen measurable. Nox publishes high-level recall and false-positive metrics for that detector on its Trust & Transparency page. Those published metrics apply to the deterministic detector; the separate AI backstop is not combined with them.
This does not mean deterministic rules catch every possible emergency. Leo is explicitly a safety net, not a guarantee. It means that a defined first layer can be tested, maintained, and evaluated independently from the conversational system that may answer an everyday question afterward.
The tradeoffs between fixed rules and AI-based screening are worth understanding, particularly in health contexts. Rule-Based Safety Systems vs AI Safety Models explores why a layered design can matter.
Why Leo is visible and controllable
Leo is not an invisible background policy. It is named, user-visible, and available from the chat composer, where people can choose how far its screening goes.
Nox offers three levels: Relaxed, Standard, and Strict. Relaxed warns about clear emergencies only. Standard also warns about situations that may be urgent. Strict adds the AI backstop that reviews recent messages for dangerous descriptions expressed indirectly or in another language.
No matter which level a person chooses, Leo always screens for true emergencies. The setting changes optional layers on top of that baseline; it does not turn off emergency screening. Agent and voice conversations use the strictest setting.
Making this control visible is part of treating safety as a product feature rather than a hidden promise. It gives people a clearer view of what the system is doing and where its boundaries are.
What happens when Leo detects an emergency?
When Leo identifies a likely emergency, Nox presents emergency or urgent-care guidance before any AI-generated response. If a user has set their region in Settings, the banner shows the appropriate local emergency number.
If no region is available, Nox falls back to universal emergency numbers so the guidance is not blank. In a real or immediate emergency, the right action is to contact local emergency services rather than rely on a chat.
Leo’s role is to interrupt the normal flow of a health conversation when a red flag may be present. It does not decide what condition someone has, and it does not provide a substitute for professional assessment. Its job is to make the safety signal visible early enough that it cannot be easily missed.
For more detail on this experience, see What Happens When Leo Detects an Emergency?.
Why safety is more than answer accuracy
A health AI system may produce an answer that sounds informed and still fail at the most important moment: recognizing when it should not continue as though the situation were routine. Accuracy and safety overlap, but they are not the same measure.
Safety includes what the system does before it answers, whether it can recognize urgent language, whether its warning is prominent, and whether it communicates appropriate next steps without diagnosing. It also includes being candid about limits.
That is why Nox built Leo as a separate medical-safety layer rather than relying only on the quality of a conversational answer. In health conversations, knowing when to interrupt can be as important as knowing how to explain.
Common questions
Is Leo a medical diagnosis tool?
No. Leo screens for acute medical red flags and surfaces guidance to seek appropriate care. It does not diagnose, treat, or prescribe.
Can Leo catch every emergency?
No. Leo is a safety net, not a guarantee. No automated system catches every emergency, which is why anyone experiencing an immediate emergency should contact local emergency services.
Does Leo work only with English messages?
No. Leo’s deterministic screening includes multilingual coverage, and Strict mode adds an AI backstop intended to help recognize indirect phrasing, slang, and descriptions in another language.
Can I see how Leo is evaluated?
Yes. Nox publishes high-level recall and false-positive metrics for Leo’s deterministic detector on its Trust & Transparency page.