A Reddit developer proposes training the pretrained Qwen 3.5 4B as a specialized coding agent, beginning with a pilot covering 100–200 functional mini-SaaS applications and tasks such as development, testing, debugging, and maintenance. The developer says supervised fine-tuning alone in earlier work produced complex Three.js scenes, voxel structures, and outputs exceeding 64K tokens, though those results are self-reported and not independently verified. The author is seeking advice on coding domains, agent harnesses, and evaluation methods before proceeding and plans to release weights, datasets, and experimental results, including unsuccessful runs.