Jev in action
Jev in action – task classifier and auto model router in Keen Code Hi HN! In Keen Code, a coding agent solo-developed by me, I have implemented a task-classifier using Jev and a simple auto-router based on the classification result. The article shares details about how it's implemented. In short, it has a very simple idea: - For every user message, ask Jev to classify the task in one of the three categories (simple, standard, complex) given current state of the conversation. - Keen also asks Jev to put a probability on this question: "is this a new task?" - Based on the result, Keen uses a heuristic to switch the current model to one of the three pre-configured models by user for the three categories - To make the router work, users have to configure one model per task category inside the coding agent itself The heuristic is also very simple: When routing is enabled, Keen switches only if the selected-category probability is greater than 0.5 and either: - new-task probability is at least 0.6; or - category probability is greater than 0.8 and the target category is a higher tier than the active routed model. Thoughts and feedback are welcome. And if you are interested to know more about Keen Code, the repo can be found here: https://github.com/mochow13/keen-code