Careline

Inspiration

We have all experienced the frustration of waiting on hold just to book a quick doctor's appointment. At the same time, medical front desks are drowning under an endless wave of repetitive phone calls. We wanted to build a solution that instantly helps patients while giving hours of productive time back to overworked healthcare staff.

What it does

CareLine is an AI voice assistant that answers clinic phones immediately. It checks the caller ID to greet returning patients by name or guides new patients through a smooth, multi-step registration process. The agent pulls live data to answer insurance and clinic questions, fills out appointment slots, and sends text confirmations. If a caller mentions a dangerous symptom like chest pain, it immediately gives a 911 advisory and transfers them to a human line.

How we built it

We connected a Twilio toll free number to Twilio ConversationRelay, which streams text transcripts back and forth over WebSockets. This let us completely bypass raw audio management. The core engine is built on LangGraph as a strict state machine to prevent the AI from losing track of the workflow. We paired Anthropic Claude Haiku for fast routing with a SQLite database for patient records, ChromaDB for clinic document lookups, and Redis to keep the entire setup fast and scalable.

Challenges we ran into

The biggest hurdle was handling conversational turn-taking over a phone call. Because voice callers cannot wait around for an entire AI workflow to finish, we had to design the server to respond instantly after every single sentence. We solved this by saving the conversation state in Redis between turns and forcing the AI to only ask one question at a time. Managing latency was also tough, but using a fast model for routine routing kept our response times under 1.5 seconds.

Accomplishments that we're proud of

We built a completely functional, safe, and personalized voice agent in just 24 hours. We are especially proud of our intake correction loop. If a caller corrects a single detail during registration, the agent only re-fills that specific slot instead of forcing the patient to start the whole form over again.

What we learned

We learned that building voice applications requires a completely different mindset than building text chatbots. Prompt engineering alone is not enough to keep a phone conversation on track. You need a rock-solid state machine to guide the flow, protect patient data, and enforce strict safety guardrails.

What's next for Careline

We want to take CareLine from a hackathon prototype to a production-ready tool. Our next steps include certifying the platform for medical data privacy compliance, integrating it with real-world Electronic Health Record software, and adding live voice language translation to support non-English speaking patients.