AI Code Reviewer

AI Code Reviewer: VS Code extension that detects, explains & fixes AI-generated code. Uses Hugging Face, Quantum Scoring & Wikipedia API. Optimize, explain, and detect, all in the click of a button.

Comment

Inspiration

Vibe coding and generative AI use in writing snippets of code have become ubiquitous in the world of tech; however, just as these have become ubiquitous, so have the plethora of errors, hallucinations, plagiarism, and imprecision that frequently accompany AI-generated code. As a result, we wanted to bridge the gap between AI-assisted coding and production-ready quality. Inspired by tools like GitHub Copilot but frustrated by their lack of transparency, we built AI Code Reviewer, a VS Code plugin that not only detects AI-generated code but actively improves it.

What it does

AI Code Reviewer combines static analysis, machine learning, and quantum computing to identify AI-generated patterns in user-inserted code. Hence, when a user inserts code into VS Code, the plugin will return an AI-likelihood score, telling the user how likely something is to be AI-generated, and show the user a dashboard that includes: AI-Detected, Confidence, Issues Found, and Suggestions for improvement. To detect AI, this plugin uses two approaches: the detection of common AI behaviour signatures, such as abrupt large code insertions, overly generic naming conventions, and chunks of inoperable code, and machine learning that integrates Hugging Face transformers to fact-check code. The latter allows our plugin to go beyond simple detection and debugging by integrating Hugging Face's CodeT5 for intelligent optimisations, transforming verbose or inefficient AI output into clean, maintainable solutions. For deeper validation, it cross-references code comments against Wikipedia’s API to catch factual inaccuracies. What truly sets it apart is its adaptive learning: user feedback fine-tunes detection models over time, reducing false positives and tailoring suggestions to a team’s coding standards.

How we built it

Prompt engineering techniques were incorporated into the coding process to build up the basic program architecture. More specifically, we integrated tools such as DeepSeek, GPT-4o, and Cursor to edit our code and construct some of the foundations for the frontend and backend. For the frontend, we used TypeScript and JavaScript to create features such as inline diagnostics, one-click fixes, feedback collection, and the auto-detection of AI-generated code. Edit history tracking is used to detect bulk pastes via document change events, an interactive UI undergirds lightbulb actions (Explain, Fix, Optimise) using CodeActionProvider, and the feedback dashboard is a webview panel that reviews past corrections. These features are buttressed by a backend logic constructed using real static analysis, integrated code optimisation with change metrics, quality scoring, and timeout protection. Our backend also incorporates multi-language support, i.e., it can work with Python, Java, or C++ code. The backend integrates FastAPI. For the AI detection, we used a combination of regex, spaCy/SBERT, and Wikipedia API. We used PennyLane to weigh AI probability and Hugging Face transformers to suggest cleaner and safer code.

Challenges we ran into

One challenge that we encountered was the difficulty in connecting the frontend with the backend. To overcome this challenge, we integrated prompt engineering techniques to streamline the debugging process while improving the PORT connections and integrating APIs. Another challenge we faced was optimising Hugging Face pipeline calls to stay under a 30-second timeout. Finally, we encountered the challenge of integrating the AI analysis with the code optimiser/analyser, i.e., making what was previously hard-coded into something dynamic.

Accomplishments that we're proud of

We are proud of being able to produce a product with so many useful features within a very short span of time. Some of these accomplishments include, but are not limited to: seamless integration (detects, explains, and fixes without disrupting workflow), Quantum-Augmented Scoring, which uniquely combines NLP with quantum probability weighting, User Feedback System, which is a local storage fallback that ensures offline usability, and the creation of an open source ready plugin for VS code.

What we learned

We learned how to integrate quantum computing into a machine-learning-based VS Code plugin, how to fluidly collaborate using Git and GitHub, and how to integrate version control.

What's next for AI Code Reviewer

We plan to further eliminate bugs, expand language compatibility, train domain-specific detectors (e.g., for fintech or embedded systems), create a custom rule engine, allowing users to highlight their own AI "red flags", and officially publish the plugin on VS Code.

Built With

Share this project:

Updates