ClinicFlow

Inspiration

ClinicFlow was inspired by how much clinic front-desk work is repetitive but still high-stakes: booking appointments, remembering returning patients, avoiding duplicate confirmations, and routing emergencies quickly. We wanted to show that AI can be more than a chatbot. It can operate a real workflow with tools, memory, state, and guardrails.

What it does

ClinicFlow AI acts as an AI clinic receptionist. Patients can chat with it to ask about availability, book or reschedule appointments, receive confirmation messages, and get remembered across sessions. It also detects urgent symptoms, avoids duplicate bookings, enforces time-of-day scheduling rules, and isolates patient sessions by phone metadata to prevent cross-patient data leaks.

How we built it

We built ClinicFlow with a React frontend, an Express/TypeScript backend, and the OpenAI Agents SDK. The receptionist agent uses structured tools such as check_doctor_availability, book_appointment, reschedule_appointment, fetch_patient_history, check_active_bookings, and flag_emergency_case. Patient memory and appointment state are stored locally for the demo, with clear contracts for replacing them with production databases, calendar systems, and messaging APIs.

Challenges we ran into

The hardest parts were making the system behave like real clinic software instead of a generic chat app. We had to add strict scheduling guardrails, prevent duplicate booking loops, handle streaming tool events, protect patient session isolation, and avoid unsafe medical behavior. We also had to fix UI layout issues so long conversations did not hide the message bar.

Accomplishments that we're proud of

We are proud that ClinicFlow has real workflow depth: streaming responses, visible tool execution, persistent patient memory, emergency routing, duplicate-booking prevention, strict time-window filtering, and multi-patient session isolation. The app feels like an operational front desk system, not just a conversational demo.

What we learned

We learned that useful AI agents need more than prompts. They need state machines, tool contracts, memory policies, retry logic, safety boundaries, and strong UI feedback. We also learned that healthcare-adjacent workflows require special care around identity, privacy, and emergency escalation.

What's next for ClinicFlow

Next, ClinicFlow could connect to real clinic infrastructure: Google Calendar or an EHR scheduling system, WhatsApp Business or SMS APIs, a production database, staff escalation queues, and role-based dashboards. We would also add authentication, audit logs, multilingual support, analytics, and human handoff for complex cases.

Built With

• openai
• openaisdk
• typescript
• vercel