TheraMatch

landing page
chatbot
multimodal input
primary result
secondary results

Inspiration

There wasn't much, originally we got the idea from the below article, it outlined how patient matching could be a potential use of LLMs. We looked for sources like this as we were interested in the gemini API category. https://www.ey.com/en_ca/insights/government-public-sector/six-ways-to-make-more-of-ai-in-canadian-health-care

What it does

TheraMatch that can be used to determine an ideal match for a patient and therapist using an LLM-based diagnostic conversation. It uses what is widely recognized to be the largest directory of therapists in North America, PsychologyToday. It works in a two step approach, first starting with the diagnostic conversation where the soon-to-be patient can talk about who they are and why they're seeking therapy. While this is happening, the orchestrator is noting possible filters based on the patient's history and interest. The chatbot continues to probe until the patient feels they've shared enough info to find a quality match. Then once they've began their search, the tool scrapes the remaining filtered therapists' profiles and using the previous chat history, it attempts to determine the ideal available candidate and the next best suggestions.

preferences, and goals

Filters North America’s largest public therapist directory (PsychologyToday) by geography, specialty, insurance, modality, etc.
Leverages Google Gemini’s advanced reasoning to semantically skim and rank candidate profiles
Delivers the single “ideal match” plus a short list of next-best options

How we built it

We built it using a Python FastAPI and Typescript Next.JS project. It uses the OpenAI python SDK to integrate with Gemini and OpenAI's Whisper models. It interacts with Psychology today through their available APIs and through accessing Psychology today web profiles.

Challenges we ran into

This was actually our second idea for the health care theme, but we spent too much time trying to host an electronic medical record software (OpenOSP Oscar), that we decided to move on from our previous concept. Once we decided on TheraMatch, There were some pitfalls with integrating the Gemini API while using the OpenAI SDK. It seems it struggled to interact with the Vercel AI SDK's useChat.

Accomplishments that we're proud of

We feel like we recovered in this hackathon after spending the first 16 hours focusing on our first concept.

What we learned

How great PsychologyToday's APIs are to use
Docker shouldn't suck but sometimes sucks to use
Perseverance and grit!

What's next for TheraMatch

Multi-location support! (Currently defaults to Toronto) Other than that we got a lot of things off of our bucket list towards the end, but will think of some more.