Heterogeneous Agentic AI System Using Fine-Tuned SLMs

In this work, a heterogeneous agentic AI framework is proposed for context-aware UI generation, which makes use of fine-tuned small language models (SLMs). In particular, the proposed framework splits tasks of orchestration and domain-specific generation between models, where Gemma 4 E2B is used for requirements elicitation and task decomposition, while Qwendean 4B is employed; a modified version of Qwen3-4B fine-tuned for domain-specific UI code generation.

A new iterative dataset creation procedure based on the combination of UI and TailwindCSS component ecosystems from ShadCN has been developed. Following several rounds of synthetic augmentation by means of DeepSeek-V3, the size of the created dataset increased from about 900 components to over 4,100 prompt-completion samples available under iamdyeus/ui-instruct-4k

SLM fine-tuning was done using LoRA approach, supplemented by rsLoRA stability technique in 16-bit floating point arithmetic. The trained model showed its capability to synthesize cohesive interfaces based on React/TypeScript UI code and ShadCN/TailwindCSS components.

A full-fledged heterogeneous pipeline based on LangGraph, consisting of orchestration, execution, and validation layers, was created. Experiments demonstrated an increase in the output quality consistency and lower inference time compared to the homogeneous GPT-LLM-based approach; in particular, about 3-5× lower latency was obtained within local section-based generation. 

Training artifacts, the agent harness, and an Electron application for local testing are available in the open-source repository iamDyeus/qwendean