Inspiration
The first thing you do when you're pregnant and something doesn't feel right is ask Google. Google tells you your baby's in danger. The second thing you do is ask ChatGPT. ChatGPT tells you it's twenty different things, and that you should go see a doctor. Neither helps. Both terrify.
The real problem isn't missing information — it's hesitation. Thousands of women delay calling their maternity team not because they don't care, but because they're afraid of being a nuisance, unsure whether their concern "counts," or simply don't know which door to knock on. For women from certain ethnic backgrounds, that hesitation is compounded by a well-documented risk of not being heard even when they do reach care.
Mira was built for those women. Not the ones who confidently ring the triage line — but the ones sitting with a quiet worry at 2am, convincing themselves it's probably nothing, and waiting.
What it does
Mira is an AI-supported maternity triage companion that turns a worried pregnant woman's hesitation into a legitimised, directed next step. It doesn't diagnose. It navigates.
The woman opens the app and selects her main concern from tappable symptom chips — reduced fetal movements, signs of pre-eclampsia (headache, vision changes, swelling), or bleeding and fluid — or describes it in her own words. Mira uses a GLM-powered AI layer to interpret her free text, then guides her through an adaptive set of clinical questions grounded in NHS and RCOG guidance, modelled on the Birmingham Symptom-specific Obstetric Triage System (BSOTS). A deterministic rules engine — never the AI — computes the urgency and routes her to the right NHS maternity track based on her symptoms, gestation, and her local NHS trust's specific thresholds: maternity triage (24/7 from 16 weeks), 999/A&E, NHS 111, or the Early Pregnancy Unit, with her trust's direct number and two nearby alternatives.
She sees one calm, plain-language result. No diagnosis. No list of scary possibilities. Just: where to go, how urgently, and who to call.
Separately, a clinician-only SBAR summary — written by the AI from her full session — is sent ahead to the receiving team, so the midwife starts the conversation already briefed. For under-heard mothers, that structured handoff is an advocate that speaks clinical language on her behalf.
How we built it
Mira is a mobile-first React web app (Vite + TypeScript) backed by a Java 17 + Spring Boot REST API. We chose a mobile-first web app over React Native for demo reliability — it runs in a browser, deploys to a shareable URL, and needs no emulator.
The AI layer uses Z.ai's GLM model via their OpenAI-compatible API at three touchpoints: (1) interpreting the patient's free-text input into a structured cluster and signal set; (2) conditioning the order and selection of clinical questions from an approved, safety-validated question bank — the AI adapts the conversation without inventing unsafe questions; and (3) generating the plain-language patient message and the SBAR clinician summary at session end.
The routing decision is made by a deterministic rules engine grounded in NHS/RCOG guidance — not the AI. The engine applies over-escalation (always returning the highest urgency reached), a mandatory red-flag sweep across all three complaint clusters, and trust-specific gestation thresholds drawn from a live NHS Trust data model. The question structure was modelled on BSOTS and RCOG Green-top Guideline No. 57.
Challenges we ran into
The hardest design challenge was the line between helpful and harmful. Our job was to build a tool that reduces anxiety without producing false reassurance, and that empowers action without claiming clinical authority.
Getting the AI/rules split right took serious thought. An LLM optimises for the most likely answer — but triage is about ruling out the dangerous case. A probabilistic model would systematically under-triage the rare-but-serious presentation. We kept the safety-critical routing decision deterministic and traceable, which required understanding how real maternity triage (BSOTS, NHS Pathways) actually works before we wrote a line of code.
Trust-aware routing added meaningful complexity: NHS trusts have different gestation thresholds for maternity triage and the EPU, different phone numbers, and different service hours. We had to build a Trust data model and patient trust-assignment layer before any routing logic could be wired — that ordering mattered.
Accomplishments that we're proud of
The dual-output design is what we're most proud of: the patient sees a calm destination, the clinician sees a structured SBAR handoff. It solves two problems at once — keeping the patient experience free of fear and diagnosis, while arming the receiving clinician with structured pre-triage information. For under-heard mothers, that handoff is advocacy.
We're also proud of the red-flag sweep. After assessing the patient's chosen complaint, Mira always screens for symptoms from the other two clusters — because the dangerous presentation is often the one she didn't think to mention. Catching that on stage, from an input that didn't flag it, is the most convincing demonstration the tool works.
And we're proud of building something that can be appropriately reassuring. The hardest thing in a health app isn't the alarm — it's the calibrated "this can be monitored at home." Without that, you haven't solved the problem. The same headache input produces two completely different outcomes depending on the answers: that's what calibrated triage looks like.
What we learned
The most important thing we learned: the problem isn't information, it's hesitation. Pregnant women don't delay because they couldn't Google it. They delay because they're afraid of overreacting, or because previous encounters have taught them their concerns won't be heard. A better information product doesn't fix that. A product that validates their instinct, takes them seriously, and makes the next step frictionless — that does.
We also learned to think carefully about where AI belongs in safety-critical software. AI is well-suited to language: interpreting lay descriptions of symptoms, selecting the most relevant next question, writing a clinician-literate handoff. It's the wrong tool for the binary safety decision. That should be deterministic, auditable, and traceable to a clinical guideline — the way NHS Pathways (the algorithm behind 111) already is.
And the BSOTS system, developed at Birmingham Women's Hospital, is a masterclass in structured clinical triage. If you're building anything in maternity care, read it first.
What's next for Project Mira
The immediate next step is clinical validation and the MHRA Software as a Medical Device (SaMD) classification pathway — Mira would need to be assessed and validated against real clinical outcomes before any deployment.
The distribution strategy is the bigger opportunity. Mira should be issued by maternity services at the booking appointment — the first antenatal visit — the way maternity notes apps already are. Installed, trusted, and connected to the right pathways before it's ever needed. Not downloaded in a panic.
Longer term: expanding question coverage across more maternity complaints, full multilingual support as a direct equity intervention for non-English-speaking mothers, and integration with wearable fetal movement monitors. The underlying architecture — adaptive triage intake, deterministic safety rules, dual patient/clinician output — also generalises. The same engine can serve other delay-prone, under-heard patient groups. Mira is the proof of concept for a broader platform.
Log in or sign up for Devpost to join the conversation.