ROUTEAI infra
TypesafeAI Jev for Assistant Routing
Ahmed Nadar@ahmednadar
This project utilizes TypesafeAI's Jev mechanism to improve the routing accuracy and speed of a family assistant's tests across multiple languages, significantly outperforming keyword matching.
SignalJev answers in a blink and never writes a word ⚡️ Beside, working on replacing LLM work with Jev comparison capabilities for #SolveTO It's fantastic for choosing city issues categories and options.
Pattern↑ Jev AI decision engine
This is so cool. I ran @typesafeai Jev on Loon, the family assistant's routing tests against 3 different languages, same 98 asks.
- Keyword matching: English 52%, French 42%, Arabic 33%. Ouch.
- Jev: English 98%, French 100%, Arabic 92%. awesomeness
Median 164 ms per decision. The whole run cost less than a cent.
That's the part that got me. Most of what an assistant asks a model is not writing. It's "which tool, which app, is this a command". Today that's a full model turn, seconds each & lots of tokens. Jev answers in a blink and never writes a word ⚡️
Beside, working on replacing LLM work with Jev comparison capabilities for #SolveTO It's fantastic for choosing city issues categories and options.