LIVE VOICE CONVERSATIONAL AI AGENT BUILDS
Conversational AI & Healthcare Intake
A Gemini prototype exploring how live back-and-forth voice AI can standardize a repetitive healthcare intake process to keep patients informed, prepared, and less anxious before they arrive.
Flying Straight.
I spent a few years working in casino management, which is a bit like working in wildlife management if gazelles smoked Marlboro Reds and frequently asked if you “knew a guy.”
One day a new GM waltzed in and announced we were going to start drug testing the entire staff. Reasonable idea, wrong industry.
I had two questions. One, what drugs were we going to test for? Two, how did he plan on operating a casino with exactly zero employees?
Casinos, bars, nightclubs, at the time we weren’t exactly wading into a talent pool of quilting enthusiasts. Everybody in the industry knew it, nobody talked about it, and drug testing would have emptied the building by Tuesday afternoon.
Some industries don’t get that luxury. Boeing tests mechanics by the thousands. Transit agencies test every driver responsible for forty tons of bus full of strangers. Freight, rail, aviation, construction, healthcare, it’s federal, and it happens in volume.
I got a firsthand look at the insane volume of calls working with an occupational healthcare client of mine. Some weeks they process waves of new hires from a single employer, all asking the same six questions in an endless cacophony of phone calls.
We decided to experiment with a live voice agent that could manage the entire pre-appointment conversation. The agent matches the caller with their employer, clarifies the process, books the appointment, offers directions, answers any questions and delivers a QR code on the way out.
Whether it’s twenty calls a day or two hundred, nobody has to explain the same thing twice.
CLEARPATH IN ACTION
Greet, schedule, guide, and check in.
All in 5 minutes or less.
Each user request streams audio in real time to a Python backend over a WebSocket connection.
Gemini processes speech natively, no speech-to-text, no text-to-speech, no translation layer in between. The model hears the caller and responds as audio directly.
A fuzzy match tool checks the caller’s employer against records on file and routes to the correct drug panel automatically. If no match is found, the agent hands off to a callback instead of guessing. As the conversation moves from employer lookup to scheduling to directions to confirmation, the screen advances in sync, nine stages, each one tied to exactly where the conversation actually is, not a script assuming where it should be.
While built around one clinic’s intake flow in this example, the same architecture extends to any process where a caller needs to be identified, scheduled, and informed before a single human picks up the phone.
Want to experience your own run of the agent? Please reach out at michaelnworden@gmail.com.
CLEARPATH TOOLKIT
- Gemini 3.1 Flash Live Preview: Native Audio-to-Audio Model
- Python ADK Backend with WebSocket Streaming
- Fuzzy Employer Lookup with Automatic Drug Panel Routing and Callback Escalation
- Tool-Grounded Scheduling with Hallucination Prevention
- Dynamic Screen Sync Across 9 Conversation Stages
- Real-Time Directions via Google Maps Integration
- Guardrails Against Cheating Guidance, False Promises, and Outcome Predictions