ToothiOM

Inspiration

Our team works in dental staffing software. It is increasingly difficult for practices to fulfill their patients needs with full time employees alone, so they often rely on temp workers to fill gaps in their schedules. Part of our goal is to reduce the burden of dental office managers day to day, so we dedicated our time at the hackathon to automate some of their workload.

In preparation for the hackathon, we spoke to a few office managers who actively use our platform about problems their practice is facing and where they're spending significant amounts of their time. All of them stated that beyond staffing, filling up their appointment schedule is the next big challenge. We believe we can leverage AI agents to automate a large part of their work.

A significant production loss for practices is Patients not scheduling follow up appointments. The data varies but generally around ~20% of patients don't attend a follow up appointment. That can happen for a variety of reasons but one that sticks out is issues due to the office staff either not scheduling the appointment or making a mistake (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1279909/). With each appointment being worth roughly ~$225 for a practice in total production that means that annually these scheduling issues are creating a $35,000-$50,000 loss of production for the practice. Across the 200,000 practices in the US that creates a nearly 10 billion dollar loss in production annually.

What it does

ToothiOM is an OM (Office Manager) agent with a set of tools that enable it to schedule patient appointments, intake appointment requests for new and existing patients, and send appointment reminders. ToothiOM interfaces with patients over voice calls.

How we built it

We built ToothiOM on top of our existing product stack at Toothio. We first intake doctors notes through our system. Then use ai agents to determine what follow up actions need to be taken / if we need to schedule a follow up appointment given what the doctor recommended. We use Telnyx to intake and send calls to patients. Then we stream the audio from the call to our server where we use a model from deepgram to translate that audio to text. Next we feed that text into a series of agents that are all equipped with different tools to interact with our system, verbally respond to the patient, and schedule the appointment.

Challenges we ran into

The first snafu we encountered was working with the Telnyx voice API and using it to live prompt GPT-4o as the patient talked. We initially struggled to keep response times down in order for the conversation to feel natural.

Accomplishments that we're proud of

We were able to get the full core flow to work in the short period of time. The system calls the user, books the appointment by having a conversation with the user to find a time that works for both the practice and the patient, and handles any scheduling conflicts automatically.

What we learned

The challenges behind making voice to voice generation feel natural with generative text models.
How to architect the agent to allow us to work in parallel and implement tools independently ## What's next for ToothiOM
Reduce latency by prebaking operations / responses
Send automatic confirmation text / follow up calls to confirm they are still attending
intake reschedule calls & new patient intake
sync the appointments with the practices management system (dentrix, eaglesoft, etc)