RepoSense-AI

RepoSense-AI is a full-stack application designed to simplify the process of understanding unfamiliar GitHub repositories by automatically generating structured summaries. Developers often spend significant time exploring file structures, dependencies, and documentation when working with new or open-source projects. RepoSense-AI addresses this problem by providing instant insights into a repository’s purpose and architecture.

The system uses a hybrid approach combining rule-based analysis and local Large Language Model (LLM) inference to ensure both accuracy and efficiency. Instead of relying solely on AI, RepoSense-AI first processes the repository structure using deterministic logic to identify important files (such as entry points and core modules) and filter out non-essential components like test, CI/CD, or configuration files. This structured input is then passed to a locally running LLM (Phi-3 via Ollama) to generate concise and meaningful summaries.

A key focus of this project is reducing LLM hallucinations, a common issue in AI-generated outputs. RepoSense-AI enforces strict constraints on the model by limiting it to the provided file structure and preventing assumptions about missing data. This results in more reliable and grounded outputs compared to typical AI-based analysis tools.

The application follows a modular full-stack architecture:

• A backend (Node.js) handles repository data processing, GitHub API interaction, and AI communication

• A service layer manages logic for GitHub data extraction and AI summarization

• A utility layer processes and structures file data for accurate analysis

• A frontend (React) provides a simple interface for users to input repository links and view generated summaries

RepoSense-AI outputs key insights including:

• Project Type

• Tech Stack

• Key Modules

• Confidence Level

• Functional Summary

This project demonstrates the effective integration of software engineering principles with AI systems, focusing on reliability, structured reasoning, and real-world usability. It is especially useful for developers, contributors, and students who want to quickly understand codebases without manually exploring every file.