Invite-only testers
MIU, ED, orthopaedic and VFC clinicians, plus junior doctors interested in reviewing safety-before-RAG behaviour, retrieval grounding and fracture-detection research.
AI / RAG portfolio · invite-only clinician testing
FractureFlow AI
A clinician-facing fracture decision-support research MVP built around safety-before-RAG retrieval, recommendation validation and X-ray detection modelling.
Designed for UK MIU, ED, orthopaedic triage and Virtual Fracture Clinic workflows. Currently open only to invited clinicians and junior doctors evaluating the prototype — not for patients or public distribution.
What FractureFlow AI is
FractureFlow AI helps trained clinicians capture structured injury information, apply deterministic red-flag checks, retrieve source-linked UK guidance, and review immobilisation / referral / safety-netting suggestions before acting.
It is clinical decision support, not autonomous diagnosis. Local Trust guidance and senior clinical judgement always override the app. Routine recommendations must cite clinician-CDS-approved sources only.
- MIU and urgent treatment centre workflows
- ED musculoskeletal triage support
- Orthopaedic triage and VFC referral
- Deterministic red-flag escalation
- Source-linked UK guidance retrieval
- Immobilisation and discharge wording aids
- Offline bundled knowledge pack
- X-ray detection research track
Who can test it now
Distribution is intentionally narrow while clinical safety, Trust localisation and evidence governance continue.
Not available to
Patients, carers, the general public, or any App Store / open distribution channel until governance sign-off and Trust-local pathways are in place.
RAG, safety & modelling phases
This page is written as an AI / modelling portfolio piece. The build order was deliberate: register evidence, lock red flags, gate CDS sources, then consider generation and imaging models.
Evidence pipeline
Source register, ingestion rounds, chunking and governance metadata across national guidance, VFC examples, red-flag packs and immobilisation material.
Safety-before-RAG
Deterministic red-flag rules run first. Escalation bypasses retrieval entirely so generation never softens an emergency pathway.
Governed retrieval
Token-overlap retrieval with authority weighting, priority boosts and a cds_only filter so routine clinician recommendations cite approved sources only.
Recommendation validation
A validator gates routine outputs and returns an insufficient-source fallback when evidence support is missing — the optional LLM cannot bypass this gate.
Evaluation harness
Shared vignette schema, Python/Flutter parity tests, RealCaseStudy de-identified cases, and clinician review workflows before any live deployment.
X-ray detection research
Separate ML track: ensemble screening (EfficientNet / SigLIP / ViT), Grad-CAM overlays, triage grading research, and double-review human labelling packages.
Runtime recommendation pipeline
Every case follows a fail-closed path. Escalation is deterministic; generation is optional and validator-gated.
01 captureStructured injury data: body region, mechanism, neurovascular status, skin integrity and red-flag prompts.
02 screenDeterministic safety rules (open fracture, NV compromise, compartment, infection, safeguarding, major trauma, and more).
03 escalateRed-flag hits return escalation prompts and bypass retrieval / generation entirely.
04 retrieveOffline / engine retrieval over the governed chunk index with CDS-only filtering for routine recommendations.
05 composeTemplate or optional LLM draft from retrieved evidence — LLM disabled by default in research builds.
06 validateRecommendation validator blocks unsupported output and forces an insufficient-source fallback.
07 presentClinician reviews source-linked suggestions, immobilisation notes, referral checklist and safety-netting wording.
08 overrideLocal Trust pathways and senior judgement remain authoritative; the app never claims autonomous diagnosis.
Modelling choices that matter for hiring conversations
The modelling story is about constraint, evaluation and clinical governance — not unconstrained chat or black-box imaging claims.
Safety before retrieval
Red flags are rule-based and run before RAG. Escalation cannot be softened by a generative model.
CDS source gating
Only a small approved subset may drive clinician disposition suggestions; patient-education material stays discharge/safety-netting only.
LLM deferred by default
The LLM provider exists but is off until tracing, approval and validation gates are strong enough for research demos.
Offline knowledge pack
Flutter ships a bundled retrieval pack so MIU/ED evaluation can run without depending on a live networked model.
Parity evaluation
Shared vignettes keep Python engine and Flutter safety logic aligned; RealCaseStudy adds de-identified ED/MIU cases.
Imaging as a separate track
X-ray ensemble research and human double-review labelling are explicit modelling workstreams, not silent auto-diagnosis.
Technical stack snapshot
| Client | Flutter (iOS + Android) — Assess, History, Sources, Settings |
|---|---|
| Safety engine | Python RecommendationEngine with deterministic red-flag screening before RAG |
| Retrieval | Token overlap + authority weight + priority boost; CDS-only filter for routine outputs |
| Corpus | ~150 registered sources / ~845 chunks (runtime pack sync tracked separately) |
| LLM | Optional provider, disabled by default, cannot bypass recommendation validation |
| Evaluation | 16 vignette harness (expanding), Flutter parity tests, 41 RealCaseStudy cases |
| X-ray ML research | Ensemble + Grad-CAM prototype; human double-review packages; review wizard at xray-review |
How invite-only testing works
- A clinician or junior doctor requests access and confirms this is research-only, not live care.
- Invited testers exercise red-flag screening, retrieval grounding, recommendation validation and optional imaging research tools on non-live scenarios.
- Feedback feeds evaluation vignettes, source-approval decisions and Trust-localisation planning.
- Patient, public or live clinical release remains blocked until governance sign-off and local pathway approval.
Frequently Asked Questions
Is FractureFlow AI available to patients?
No. The app is not for patient self-triage or public distribution. Access is limited to invited clinicians and junior doctors evaluating the research MVP.
What AI techniques does FractureFlow AI use?
It uses deterministic red-flag escalation before retrieval, offline governed RAG over UK guidance chunks, recommendation validation, an optional LLM provider disabled by default, evaluation vignettes, and a separate X-ray detection research track with ensemble models and Grad-CAM.
Does the app diagnose fractures autonomously?
No. FractureFlow AI is clinical decision support. It does not replace clinician judgement, local Trust pathways, or radiology reporting.
Who can request access?
MIU, ED, orthopaedic triage and Virtual Fracture Clinic clinicians, plus junior doctors interested in testing the safety-before-RAG and fracture-detection research prototype, may request invite-only access.
Is this affiliated with the NHS?
No. FractureFlow AI is an independent research MVP. It is not affiliated with, endorsed by, or approved by the NHS, NICE, BOAST, or any hospital trust.
Why show this on a public website if it is invite-only?
The public page documents the AI / RAG / modelling approach for portfolio and collaboration purposes. The product binary and clinical distribution remain closed.
Important notes
FractureFlow AI is an independent research MVP by Random Mini Apps. It is not affiliated with, endorsed by, or approved by the NHS, NICE, BOAST, MHRA, Apple, Google, or any hospital trust.
Example pathways are not Trust-local protocols. Do not use this prototype for live patient care or patient self-triage. In an emergency, follow local escalation pathways and call 999 when required.
Request clinician testing access
If you are a clinician or junior doctor interested in evaluating the safety-before-RAG / fracture-detection research prototype, get in touch. Patient and public distribution remain closed.