The company says the agent asks for a budget before paid jobs, runs a small test before a full run, then trains, evaluates, and publishes outputs to the Hub.
The post reports six projects with total CPU and GPU job charges of about $103. One Qwen3.5-2B citrus-disease model was fine-tuned on 3,017 annotated images: reported accuracy on 335 test photos rose from 14.9% for the base model to 52.8% after two epochs on one A10G, at about $1.90. A 0.8B prompt rewriter was trained and released as an 812 MB GGUF file for CPU use; the reported project cost was about $16.
Other examples include Qwen-Image 2.1 LoRAs, a FLUX.2 klein LoRA, and a four-step distilled image model. The source recommends specifying a baseline, a smoke test, expected deliverables, and a spending cap. It does not establish that the reported costs, quality results, or agent reliability will generalize to other datasets, models, or teams.