Perian
My framework for letting small AI models on my own computer do the daily work for free. A frontier model designs the prompts and tests once; local models then run them every day, and a model from a different family checks every score.
- Why
- Most of what an agent does day to day, like pulling out details, scoring and ranking, doesn't need an expensive model. I wanted that work private, on my laptop, at no cost per run, and still trustworthy.
- How long
- April to July 2026
How I built it
- I split the work in two: a frontier model plans once, small local models do the daily runs.
- I had the frontier model design the prompts and an evaluation for each job.
- I gave every model a six-level test battery so models audition for roles instead of chasing one leaderboard.
- I paired a Qwen-family pipeline with a Gemma-family challenger that re-checks each score and cites the rule broken.
- I added a refine loop: a frontier model suggests prompt changes offline, and a fixed gate keeps them only if the measured numbers improve.
- I published a working job-search example that runs on placeholder data.
- To copy it you need
- A computer that can run local models, Ollama, and Claude, or the AI of your choice, for the design work. Gotcha: a second opinion from the same model family tells you little, because they share most of their mistakes. Use a different family to check.
Try it yourself
Copy this prompt and paste it into ChatGPT or Claude to try it yourself.
I want to run a repetitive AI job on my own computer for free instead of paying per request. Korab Eland's Perian framework (on whatarewebuilding.ai, code at github.com/korabeland/perian) does this: a frontier model designs the prompt and a test once, a small local model does the daily work through Ollama, and a model from a different family re-checks every score, because models from the same family share most of their mistakes. Help me set up a small version for one job I do often. First ask what the job is and what computer I have. Then pick two local models from different families, write the prompt, and write ten test cases so we can measure it. Ask me for anything you need.
Tools I used for this project
