Architecture | Clinical Coverage | Why This Matters | Roadmap | Tech Stack | Team
Every GP in the NHS gives the same instruction: "Come back if you're not better." Most patients don't come back. They deteriorate at home, misread their own symptoms, or simply forget. Safety netting - the clinical process of ensuring patients know when to seek further help - is verbal, untracked, and fails silently.
SafetyNett closes that gap. GPs create condition-specific safety nets at the point of care. The system contacts patients automatically after a set interval, collects their response, and uses AI to analyse whether their symptoms indicate clinical deterioration. If the AI detects red flags, the GP is escalated immediately with the patient's own words and a severity assessment.
Live: safetynett.lovable.app
Built in 2.5 hours at the OpenClaw Clinical Hackathon (28 March 2026, Accurx HQ, London).
Hackathon prototype, not a deployed clinical device.
GP creates safety net ─→ Timer expires ─→ Patient contacted (email / voice)
│
Patient responds
│
AI analyses response
against condition-specific
red flag criteria
│
┌───────────────┼───────────────┐
▼ ▼ ▼
No red flags Red flags No response
→ Resolved detected → Escalate
→ Severity (non-response
scored is clinical
→ GP escalated signal)
The prototype prompts the model to evaluate the patient's natural-language response against condition-specific red flags. A parent writing "she won't drink anything and she's really floppy" should trigger escalation because fluid refusal and reduced consciousness are clinically concerning in a febrile child. This behaviour still needs formal clinical safety validation before real deployment.
39 conditions across 9 specialties. Red flag definitions sourced from NICE Clinical Knowledge Summaries and NHS 111 pathway documentation.
| Specialty | Conditions |
|---|---|
| Paediatrics | Viral URTI, bronchiolitis, croup, febrile convulsion, meningitis |
| Respiratory | CAP, asthma exacerbation, PE, COPD exacerbation, suspected lung malignancy |
| Cardiology | ACS, heart failure, AF, pericarditis, aortic dissection |
| Gastroenterology | Appendicitis, pancreatitis, upper GI bleed, IBD flare |
| Neurology | Concussion, migraine, SAH, first seizure, TIA |
| ENT | Tonsillitis, peritonsillar abscess, otitis media, epiglottitis |
| Mental Health | Acute suicidal ideation, psychotic episode, acute anxiety |
| Musculoskeletal | Suspected fracture, septic arthritis, cauda equina syndrome |
| Dermatology | Cellulitis, allergic reaction, necrotising fasciitis |
Each condition carries its own red flag set. Meningitis flags non-blanching rash, neck stiffness, photophobia, bulging fontanelle, altered consciousness. Cauda equina flags saddle numbness, bilateral leg weakness, urinary retention. The clinical logic is specific, not generic.
Safety netting is an established clinical concept with weak digital infrastructure. NICE guidelines mandate it. The RCGP teaches it. Every GP does it verbally. Existing tools (EMIS templates, Ardens, SystmOne flags) are passive reminders - they don't contact the patient, collect a response, or analyse whether symptoms have progressed. The gap this prototype targets is active follow-up after the patient leaves the consultation.
This isn't a theoretical gap. NHS England stated in October 2024 that primary care should "have information systems that automatically flag patient safety issues such as missed patient referral follow-ups, safeguarding, diagnoses and medication issues." A 2022 JMIR framework for evaluating e-safety-netting tools concluded that no tools currently available meet all the criteria. Existing solutions (EMIS templates, Ardens, SystmOne flags) are passive reminders embedded in EHR systems - none actively contact the patient, collect a response, or analyse it.
SafetyNett is not a symptom checker. It doesn't diagnose. It automates the follow-up that GPs already give, makes it trackable, and escalates when patients describe symptoms that clinically warrant it.
The primary feedback from hackathon judges (NHS clinicians): automated systems contacting patients about clinical symptoms in a regulated environment carry risk. The right response is not removing AI - it's adding a verification layer.
A dedicated verification model reviews every outbound communication and every escalation decision before it reaches a patient or GP. The verifier checks:
- Whether the AI's clinical reasoning is sound for the specific condition
- Whether severity classification matches the described symptoms
- Whether edge cases (ambiguous language, multiple conditions, patient distress) are flagged for human review
This creates a dual-model pipeline: primary AI analyses the response, verification AI audits the analysis, human clinician intervenes where either model flags uncertainty. Designed for MHRA Software as a Medical Device (SaMD) classification and NHS clinical safety standards (DCB0129 / DCB0160).
Email excludes the patients who need safety netting most - older patients, patients with visual impairments, patients with low digital literacy. These are the same patients most likely to deteriorate without seeking help.
ElevenLabs voice synthesis delivers follow-ups as automated phone calls running simultaneously with email. The patient responds verbally; speech-to-text feeds into the same AI analysis pipeline. Same clinical logic, different channel. The GP's dashboard shows both email and voice responses in one view.
- FHIR integration for NHS Spine connectivity (PDS patient lookup, GP Connect)
- Structured audit logging for CQC inspection readiness
- Multi-language patient communications
- GP-configurable custom red flag definitions
- Frontend: React, TypeScript, Tailwind CSS, shadcn/ui
- Backend: Supabase (PostgreSQL, auth, edge functions)
- Email: OpenMail API
- Deployment: Lovable Cloud
src/
├── components/ Dashboard, safety net cards, forms, navigation
├── hooks/ Auth, toast, mobile detection
├── integrations/ Supabase client and types
├── lib/ Clinical logic - conditions, red flags, timeframes
├── pages/ Dashboard, create safety net, login
supabase/
├── migrations/ Database schema
├── functions/ Edge functions (trigger-followup, process-response)
The app runs at safetynett.lovable.app. All backend logic (auth, database, edge functions, email dispatch) runs on Supabase Cloud - there is no self-hosted mode.
For local development (modifying the frontend):
npm install
npm run devRequires .env with Supabase and OpenMail credentials pointing to the cloud instance - see .env.example.
npm run lint
npm run typecheck
npm run build
npm testGitHub Actions runs the same checks on every push and pull request.
Hackathon prototype. Supports clinical decision-making - does not replace clinical judgement. AI verification layer and regulatory compliance work are in active development.
Built at the OpenClaw Clinical Hackathon by Ekin Kahraman and Simon Yip, as part of a five-person team.
Clinical logic and safety-netting criteria were authored from NICE Clinical Knowledge Summaries and NHS 111 pathway documentation, with clinical review by Simon Yip.
MIT