AlfaCura

Italian patients looking for a physiotherapist today pick one from a directory, hope the specialization matches, and often discover mid-visit that the physio doesn't actually treat their condition. Physios on the other side waste calendar slots on non-targeted patients, field the same WhatsApp questions all day, and juggle bookings, invoices, consent forms and reviews across five disconnected tools.

The brief was to build a single product that:

• Collects a patient's symptoms in conversational Italian, safely.
• Matches them to the three best-fit physios (specialization, zone, availability — not just distance).
• Gives the physio a management suite that replaces the five tools.
• Stays compliant with Italian healthcare data rules — PHI encrypted at rest, signed consent, audit trail.

1. The triage bot — 5-layer Italian clinical pipeline

triage_bot.py (~2,800 LOC, Python/FastAPI) replaces a naive "send the message to the LLM" approach with a deterministic pipeline:

• Layer 1 — Safety detector. Hardcoded regex for cardiac, stroke, self-harm and other emergency phrasings in Italian. Catches the cases an LLM must never fumble.
• Layer 2 — Intent router. Rules-first, LLM fallback. Distinguishes data_response, question, crisis_signal, off_topic so the dialog manager knows what to do next.
• Layer 3 — Entity extraction. Italian-language NER for age, temperature, pain scale, duration, body part, gender, contextual preferences.
• Layer 4 — Dialog manager. State machine with 9 red-flag investigation protocols. Stateful across turns via a dialog_state column persisted in Postgres — the conversation survives refreshes.
• Layer 5 — LLM response generator. Anthropic Claude with a tight system prompt, only invoked once the previous layers have structured the context.

The pipeline was originally written as Deno Edge Functions, then migrated to Python/FastAPI on subject-matter advice that Python would give better clinical reasoning. Frontend has a single env-var switch (VITE_PYTHON_BOT_URL) and backward-compatible endpoints at /functions/v1/claude-triage — so either backend can serve at any time and rollback is instant.

2. The matching algorithm

After triage extracts the clinical picture, an availability scanner (availability_scanner.py) walks every professional's configured working_hours, subtracts booked appointments and blocked_slots, and ranks the intersection of:

• Clinical specialization — filter by pathology tags, not just a free-text specialty field.
• Service area — only physios whose declared zones cover the patient's address.
• Real-time availability — slots calendar-synced in real time.
• Review quality — only feedback from verified visits (review rows tied to completed appointment IDs).

Output: the top 3 physios with a match percentage the patient can understand, plus an AI-generated explanation of why each is the right fit.

3. The professional suite (web + mobile)

26 screens on mobile alone, fully mirrored on the web dashboard:

4. The patient suite

A medical archive that does more than store PDFs:

• Uploads (referti, analisi, documents) auto-routed through AI and filed into ordered sections.
• History travels with the patient across physios.
• Appointments, consent forms, messages and the AI-triage chat all live in one timeline.

5. Security & compliance

Italian healthcare data is PHI. Migration 032 (phi_encryption.sql) + 033 (auto_encrypt_triggers.sql) + 034 (drop_plaintext_columns.sql) migrated the entire patient dataset to encrypted-at-rest PHI columns via Postgres triggers, then dropped the plaintext columns in the same deploy. Audit trail (audit_trail.sql) logs every read/write on a patient record. RLS policies enforce that a physio can only see their own patients.

6. Mobile app (React Native + Expo Router)

The mobile experience isn't a thin web wrapper — it's a first-class React Native app with:

• Auth-first flow with a guest mode: unauthenticated users can use triage and search; gated areas render an inline "Accedi / Registrati" card instead of a hostile redirect.
• Contextual drawer: the left drawer adapts to the current section — on Agenda it shows Oggi/Settimana/Mese toggles + hour config; on Pazienti it shows filters — while the bottom tab bar handles top-level navigation.
• Agenda config: auto-creates the professional record on first use (fixing a "profile not configured" dead end), persists weekly working hours to a JSONB column, manages blocked slots against the same table the matcher reads from.

┌─────────────────────────────────────────────────────────────────────┐
│                            CLIENTS                                  │
│  Web (React + Vite)          Mobile (RN + Expo Router)              │
│  · LandingPage               · Professional suite (26 screens)      │
│  · Professional dashboard    · Patient suite                        │
│  · Patient dashboard         · Guest mode                           │
│  · Triage chatbot            · Triage chatbot                       │
└──────────────────────┬──────────────────────────┬───────────────────┘
                       │                          │
                       ▼                          ▼
┌───────────────────────────────────┐  ┌─────────────────────────────┐
│  Python Bot Service (FastAPI)     │  │  Supabase Edge Functions    │
│  · triage_bot.py (5-layer)        │  │  · Stripe checkout/webhook  │
│  · professional_chat_bot.py       │  │  · Invoice PDF generator    │
│  · availability_scanner.py        │  │  · Notifications/reminders  │
│  · Anthropic Claude client        │  │  · Consent signing          │
└──────────────────┬────────────────┘  └──────────────┬──────────────┘
                   │                                  │
                   ▼                                  ▼
┌─────────────────────────────────────────────────────────────────────┐
│                        Supabase (Postgres)                          │
│  32 migrations · RLS on every table · PHI-encrypted columns         │
│  · professionals · patients · appointments · messages               │
│  · services · blocked_slots · invoices · consent_documents          │
│  · reviews · notifications · audit_trail · patient_chats            │
└─────────────────────────────────────────────────────────────────────┘

• 13,000+ production conversations across two languages, every one gated by deterministic safety rules before the model was invoked.
• One product replaced five tools for the physio (directory, calendar, CRM, invoicing, consent software).
• Matching quality up, filler visits down. Physios see only patients targeted to their specializations.
• 24/7 responsiveness without hiring a receptionist — each physio's chatbot answers common patient questions on their behalf.
• Compliant by construction — PHI encrypted at rest, signed consent, audit trail, RLS.

• Don't let the LLM near untrusted input unsupervised. Layering deterministic guards in front turns "AI chatbot" from a liability into a feature.
• Migrations are a product feature. The PHI-encryption rollout was three stitched migrations with triggers — deployed with zero plaintext ever reaching the final schema.
• Two clients, one contract. Keeping the web and mobile apps on the same Supabase tables (not separate microservices) made feature parity free and bug-fixes a single PR.