PromptFlow is a dual-agent AI system designed to evaluate whether automated prompt refinement improves the quality of responses generated by large language models.
The system operates through a two-stage pipeline. Agent 1 is a fine-tuned Gemma 3 1B model trained to transform raw, unstructured user inputs into professionally structured prompts using the RISE framework (Role, Instruction, Steps, Expectation). Agent 2 is a Gemma 4 E2B model that receives the refined prompt and generates the final response.
By comparing responses produced through the dual-agent pipeline against responses generated from direct raw input, PromptFlow provides measurable evidence of how prompt quality affects response depth, structure, and relevance across diverse query categories.
Built with FastAPI and React, the system runs fully locally without dependency on external AI APIs and is deployed on HuggingFace Spaces for public access.
This project was developed as a Final Year Research Project with an accompanying research paper evaluating pipeline performance across 265 test queries spanning ethical dilemmas, critical analysis, professional writing, summarization, and mathematical reasoning.