FlowCare AI – Hospital Workflow Intelligence

Inspiration

It started from a simple, frustrating pattern we kept hearing about from hospital staff: a surgery gets delayed, and by the time anyone figures out why, the patient has already been waiting for an hour. A patient gets admitted, consent is pending, instruments are ready, but the anaesthetist isn't — and none of that information lives in one place. Existing hospital software tells you that something went wrong. Almost nothing tells you why, in real time, in a way a busy OT manager can act on.

We wanted to build something that sits on top of the systems hospitals already have — instead of asking them to rip and replace an HMS they've already paid for — and actually explains what's happening across departments.

What it does

FlowCare AI is an AI-powered Hospital Workflow Intelligence & Operational Digital Twin platform. It doesn't manage patient records — it manages the flow of a patient, an instrument, and a care team through the hospital, and explains delays as they happen.

Workflow Correlation Engine — connects events across Admissions, CSSD, OT, and Recovery instead of treating each department as an island.
Explainable Delay Engine — instead of "Surgery delayed 45 min," it breaks that down: 18 min late transfer, 12 min consent delay, 8 min instrument gap, 7 min cleaning delay.
Surgical Readiness Score — a live percentage score per surgery (consent , blood , instruments , anaesthesia ) so staff can intervene before a delay happens, not after.
Digital Twin Dashboard — a live, single-screen view of patients, OT status, CSSD, and recovery.
AI Hospital Copilot — a natural-language assistant OT managers can ask things like "Why is OT-2 idle?"

How we built it

Frontend: React, TypeScript, Tailwind CSS, shadcn/ui, React Flow (for the workflow/digital-twin views), Recharts (analytics)
Backend: FastAPI (Python), REST + WebSocket for live status updates
AI/ML layer: scikit-learn / XGBoost for delay prediction, with the architecture designed to extend to computer vision (OT occupancy) and NLP (the Copilot) in later phases
Data: PostgreSQL for records, Redis for live state, TimescaleDB for time-series workflow events
Infra: Docker for local orchestration

For the hackathon build, we focused our working demo on one thin vertical slice — OT + CSSD readiness tracking and delay-cause breakdown — driven by simulated real-time events, rather than trying to build every module listed on our roadmap.

Challenges we ran into

Scoping down. Our original concept touched ten hospital modules — we had to cut hard and ship one believable, demoable slice (OT/CSSD readiness + explainable delays) instead of a shallow version of everything.
Simulating "live" hospital data realistically, without access to a real hospital's event stream, so the delay-prediction logic had something meaningful to react to.
Designing the Digital Twin view so it reads at a glance — hospital administrators don't have time to interpret a cluttered dashboard mid-shift.

Accomplishments that we're proud of

A working explainable-delay breakdown, not just a black-box "delayed" flag.
A UI that a non-technical OT manager could plausibly use without training.
An architecture that plugs into a hospital's existing systems instead of demanding they migrate off their current HMS.

What we learned

Healthcare workflow problems are less about missing data and more about disconnected data — the hard problem is correlation, not collection.
Explainability matters as much as prediction in a clinical operations context; a number without a reason doesn't change staff behavior.

What's next for FlowCare AI

Phase 2: CCTV-based computer vision for OT occupancy, HL7/FHIR integration with real hospital systems, WhatsApp/SMS alerting, predictive staffing.
Phase 3: Multi-hospital analytics, "what-if" digital twin simulation (e.g. "what if we add another OT?"), voice-enabled Copilot.
Pilot conversations with small and mid-sized hospitals, where enterprise HIS/digital-twin tooling is currently priced out of reach.