Contents

• Motivation
• Introducing ZenAI
• Data
• Model training
• Model inference
• Model testing
• User experience
• Key risks & future work
• Meet our team

Motivation

 

We built ZenAI to offer Sam and others in her shoes better access to mental health support.

Overview of ZenAI

Data

Data and code for preparing the data code be found here.

Model training

Model training code can be found here.

Model inference

Model inference code can be found here.

Model testing

Overall, our initial testing has been focused on the following:

• Key quantitative metrics: Our model outperforms Llama 2 on key LLM metrics.
• Intent detection evaluation: We tested our intent-detection classifier against 30 labeled prompts. The results are shown below and highlight that our model performs quite well. Supporting code can be found here.
• Qualitative testing: We conducted extensive user testing of the chatbot. Snippets of Zen's responses are shown above and in our UX demo. We plan to roll out our testing to other target users and therapeutic experts.

User experience

The UI/UX code can be found here.

A video demonstration of our UX can be found here.

Key risks & future work

Below is the summary of the key risks of ZenAI and enhancements that we'd like to consider for the future.

Meet our team

• Manish Kumar (LinkedIn)
• Deepak Vanjani (LinkedIn)
• Puneet Modi (LinkedIn)
• Danil Meresenschi (LinkedIn)