1 comment

[ 2.9 ms ] story [ 17.7 ms ] thread
Architecture uses GPT as a thin layer that converts user prompts into a json, which is then routed to a couple of backends, one for preflop and another for postflop situations.

There's around 5 to 10 system prompts running per message, each has the job of parsing between 1 to 3 variables.

All the backend is in python, using almost no dependencies. Frontend is pure HTML, CSS, JS (and it shows).

AMA