DecenTalk

Inspiration

Solutions such as chatbot companions\therapists have been on the scene since 2017. The problem? The are really not good... Even the most popular mental health chatbots on the market:

1. Lack the ability to respond to any input from the user
2. Train their AI models with user-collected data from their app only
3. Do not have a large enough training databases to model the complex reality of human mental health

We wanted to create a mental health companion app that uses a wide variety of therapist-patient conversations and, as such, mimics more accurately the response a patient would actually receive from a human being. All that, without asking therapists to provide us their sensitive patient conversations as training data. You can find below how we actually did this.

Our business model

The business plan behind DecenTalk is simple:

• Therapists record their patient consultations with a voice-recorder
• They run Disco on their local machine, which allows them to train an NLP model without sharing their sensitive consultation data. They can perform this even with a laptop/PC, using smaller batches
• Disco aggregates the training gradients of many practitioners, and, as such, optimizes one large chatbot model: more reliably than any single company can by collecting data manually
• Users can use our chat companion instead of waiting for 3 months to get an appointment
• In return for that, we ask for a modest fee which will serve as a payment incentive to the therapists: who put effort into recording their conversations

Accomplishments that we're proud of

1. We created our own NLP training task on Disco
2. We successfully ran decentralized training on Disco
3. We created a working chatbot, which we called Soul :), that you can interact with in real-time and actually receive decent advice on mental health issues
4. We created a service that automatically transcribes video consultations, using Google Could

Built With

• disco
• docker
• python
• react
• tensorflow