Sonia aims to build a safe AI therapist that is clinically as effective as the top 99th percentile human therapist and to make mental health support accessible to anyone.
Requirements
- Prompt Engineering
- Python
Responsibilities
- We have >50 prompts that run simultaneously throughout a conversation that need to interact with each other in realtime.
- We have prompts that run in parallel to the session and do various semantic checks (e.g. should a certain intervention be applied, is the client engaged, is there something the AI therapist has missed, is there an open topic that should be circled back to, etc etc etc).
- Some are latency critical, some can take more time to reason.
- Some need a lot of context from previous sessions, some only need the ongoing transcript.
- Some will generate an output that will be injected as “advice” into the response prompt, some will change the value of a parameter in the backend that gets used to sample other prompt inputs probabilistically
- When the client sends a message, we run several safety checks in parallel to generating a response, such as a risk detection prompt that looks out for 5 categories of risk (suicidal ideation, medical emergencies, etc).
- Upon trigger, the original response call gets aborted and instead the response model receives an injection and special instructions on how to address this potential crisis and what information it needs to collect in order to know whether the conversation can be continued or should be escalated
Other
- US citizen/visa only
- Any (new grads ok)
