| Hey everyone! Sharing this project we’ve been working on. We run a coaching app for new & aspiring managers (mostly engineering managers and PMs). Users get personalized 2-minute daily tips (we shared it here way back in 2020), and they can reach out to a coach via Intercom chat if they need help with a specific problem. We typically need at least ~15 mins or so to respond to these coaching requests via chat, and our team isn’t big enough yet to provide 24/7 coverage, so we decided to try out OpenAI’s API to see if we could build an “assistant coach” who can chime in faster and still provide actual value - partly inspired by the Marc Randolph’s (Netflix Co-Founder) quote that “the next wave of disruptive companies will be built by layering domain expertise over GPT-3”. **We used prompts to give GPT-3 context on its coaching mission, including a coaching framework & questions, we gave it access to our content library of 2-minute tips for managers, and did our best to give it a bit more personality. The results seem pretty good so far! We’ve introduced “Bunchee” to users in our app already, and just today released the web demo. There are definitely still some issues though: So far, Bunchee is good at: - empathy
- humour
- improvising (most of the times)
- understanding that she is a colleague of human coaches at Bunch What she’s not so great at: - Actually being available 24/7 (due to OpenAI API dependency)
- remembering past conversations to build more context about the user
- switching modes on the fly: coaching vs. advice vs. search vs. customer support Curious to hear what you all think! |