SC-Oracare

SC-Oracare is a modular monolith: one Django deployment on PostgreSQL and Redis, split into apps by bounded context, such as patients, referrals, clinical, revenue, operations and the patient portal. One process serves both HTTP and WebSockets through Django Channels.

Each new capability arrived as a new module with its own tables, behind a feature flag, and the referral workflow it started from was preserved through every phase; the referral core itself is never flagged. Only one part has been split out as a separate service so far, the AI capabilities, which became Oracare AI.

The diagram above is a simplified view. It shows the clinician and patient sides as separate trust boundaries because they use separate identity systems, even though they run in the same deployment.

What each person gets

• Doctors and specialists chart patients, build treatment plans, create and receive referrals, and work each referral together in a treatment room.
• Office accounts run the practice: patient registration, appointment requests, provider availability and closures, billing and accounts receivable, inventory, staff and tasks.
• Staff administrators use a separate, staff-only admin portal for user management, subscription plans, sign-in logs, compliance checks and system status.
• Patients use their own portal for their treatment plan, health record, appointments, forms with e-signature, documents, secure messages, payments and telehealth.

Referrals at the core

A referral starts in a wizard that walks through the patient, the referral, the schedule, consent and a review, and submits once, at the end. The same engine serves every entry point: a doctor, an office acting for a doctor, and an anonymous QR code link. The patient consents through a time-limited, single-use link or an encrypted QR code, with a signature.

Every referral gets a treatment room: the care team, an ordered treatment plan the team can reorder, attachments and live chat. Clinical planning feeds the network too. A treatment plan item that needs a specialist becomes a referral on its own, and the referral completes when every step of its plan is done. That status is derived from the steps, so no client can mark a referral finished while work is outstanding.

AI-assisted referral

A clinician can dictate a referral instead of typing it. The audio is transcribed, the details are extracted into a structured referral, and the wizard fills every step up to the review and stops there. Anything missing becomes a short follow-up question, answered by voice or by typing, and nothing is submitted automatically. A photographed or uploaded document can propose patient details the same way, shown for confirmation before anything is filled.

These capabilities are served by Oracare AI. SC-Oracare calls it from the server with the signed-in clinician’s delegated token, so the browser never talks to the AI service and SC-Oracare holds no model and no provider key. The audit trail stays in SC-Oracare, which remains the record of who accessed what.

Key decisions

Security and identity

There are two identity systems, and they never share a session.

Clinicians, office staff and administrators sign in through Microsoft Entra External ID. The tokens stay on the server, in the session, so page scripts can never read them, and the ID token’s expiry is the real end of a session.

Patients have standalone accounts in the portal, with their own passwords, emailed one-time codes, optional authenticator app MFA and lockout. Signing in has two halves. The first proves control of a mailbox; the second picks whose record to open, since one guardian can hold several family members, and proves that person’s date of birth. Switching to another family member drops the verification.

Authorization is layered. One shared scope rule decides which patients a clinician may see: their own, or their offices’. Clinical pages add a care team check, derived from the referral itself, so a specialist brought in on a patient can open that chart without the patient appearing in their own directory. The admin portal is gated twice, on each view and in middleware, and the REST API requires authentication by default, so a new endpoint that forgets to declare a policy fails closed.

Protecting patient data

SC-Oracare is built to HIPAA requirements. A patient’s name, date of birth, email and phone number are encrypted field by field. Because that encryption gives a different result each time, patients are matched through a keyed blind index and looked up by an unencrypted medical record number, and every patient is created through one registration path that reuses a record only on an exact match.

Nothing clinical or financial is deleted. Clinical facts retire by status, signed notes refuse edits and take addenda instead, financial rows are voided, and merged patients are retired rather than removed, because “no allergy recorded” and “allergy deleted” must never look alike in a medical record. Every read and write of patient data leaves an audit entry that records who did what to which patient, and never the clinical content.

Delivery and operations

Treatment room chat, direct messages, document capture in the wizard and live notifications run over WebSockets through Django Channels, on a Redis channel layer so any worker can reach any connection. Each person’s notifications go to their own group, and the notification is written to the database before it is pushed, so an offline user or a dropped connection loses nothing.

Background work runs on Celery with Redis: appointment reminders, recall notices, token cleanup, nightly backups and a retention purge that never touches audit records or patient data. Email goes through one gateway that queues delivery and logs without patient data.

The application ships as a Docker image and is configured entirely from its environment. Moving the live demonstration off Azure onto free-tier hosting changed settings, not code: no view, model or template changed.

Result

SC-Oracare runs today as a live demonstration on synthetic patient charts only. It has no customers, so there are no usage numbers to report; the running system is the evidence. Next is the clinician’s view of retrieval: asking a question about one patient and getting an answer grounded in, and citing, that patient’s records, through Oracare AI.

Demo

Open the live demo (opens in new tab). It holds synthetic data only and runs on a free tier, so the first load can take up to a minute.