netops RL env

autonomous reinforcement learning environment to build custom ML model for your infrastructure (optimized for your infrastructure) costs 1/100 of the frontier, and can deliver better results if finetuned with enough resources. your start could be , hey I want my model to be better at detecting ospf failures. then this project kicks off multiple things like 1/ goes to internet scrapes all the data it can on ospf failures like neighbour ship faililng, ospf state block at xstart or .. everything it can. 1.1/ these data will be turned into questions and valuable training info using layered models (cheaper ones draft and costly ones verify - modify/reject/accept saved 90% expensive model tokens by this method 2/ it spins up bunch of workers (with container labs environments) and the mother model (here the opus 5.5) plugs in some faults and sends some agents to figure it out and once it claims its done its going to be verified using some code and pipelines (part of testbed) 3/ if it passes or fails its going to go into data base 4/ now you can select a new model or a previous pre trained model and the small model is going to be trained on what do to (from the successful runs) and also what not to do (from the ones that got it wrong) 5/ evals: and once the small model is finetuned you can run evals against it and choose to again reinforce on the things it failed specifically focusing on the failure material, the same pipeline runs again, opus chooses which things to train on , collects the data spins up the env converts to training data and reinforces this loop can keep going until youre satisfied. --- did I do anything before this week? nope, the ideation was on the day this hackathon started with me and claude (voice mode) fighting over which feature goes in and which doesnt I started this machine on a new server install (old hardware) I had lying around https://drive.google.com/file/d/1ll0UCq1MoA380SwGX29wkTBdkPdYZynw/view?usp=sharing this is the picture of my claude usage in the last week you can see I started on the same day (I know this is not a valid metric but I think I produced a pretty decent output quickly thanks to AI :) ) ---------------------- IMPORTANT : to check the live demo link : https://mghack.dheerajgajula.com login : judge password : metaghack this has the readonly outputs of my entire runs you should be able to cross link with the code