Revue | GenAI hackathon 2026

A.I PR Reviewer.

Comment

Inspiration##

Hearing Friends land internships excites me, as someone who is having a hard time breaking into the field and thus, a lack on an internship I begin asking questions. I've put most of my friends through some form of interview hell to learn as much I could about how they did it and what they do.

Consistently one issue always arose; how boring it can be waiting for work to be reviewed. See as an intern (from what I understand) there isn't much expected of you, therefore making a mistake is no big deal and it can be pretty easy to complete tasks. however, That also means our work can be left in the corner of a senior devs repo, and it doesn't change how busy a senior dev could be.

Revue is designed as a tool by senior devs to organize and decrease the backlog of pull requests, so that interns like us experience the downstream effects of an efficient workflow; more reviews, more work, means more time to shine and really learn from our internships.

What it does

Revue AI-augmented GitHub Pull Request reviewer.

  • An organized space Revue provides an organized space for all PR requests.

  • Seamless Authentication: Users log in with their GitHub account, granting Revue secure, scoped access to their repositories.

  • Instant Insights: It fetches open PRs and feeds the raw diffs to a powerful GenAI model, generating an Impact Score, a concise summary, and suggested code for next moves.

  • The "Big Room" Editor: When the AI suggests a code fix, Revue provides an interactive workspace to tweak the AI's explanation and suggested code. For even more accurate diff comments.

  • Native GitHub Integration: Once the suggestion is refined, Revue posts it directly to the GitHub PR comment thread on behalf of the logged-in user.

How we built it

  • Frontend/Backend: Next.js (App Router) with React, styled using Tailwind CSS for our custom gold-and-dark-theme UI, Google Gemini visualization.

  • Database & Auth: Prisma with PostgreSQL stores our review states, while NextAuth handles the GitHub OAuth flow, ensuring we capture and securely utilize user-specific access tokens.

  • GitHub API: Utilized octokit to fetch live PR data, extract raw diff media types, and push markdown-formatted comments back to repositories.

  • GenAI Engine: Integrated the OpenAI SDK pointing to a HuggingFace-hosted gpt-oss-120b model. We engineered strict system prompts to force the model to output structured JSON analyses of code diffs. By also using Generative A.I through out the project I did what used to take months in only 30 or so hours.

Challenges I ran into

The model would occasionally wrap responses in markdown backticks or add conversational filler, crashing the UI parsers.

I had to engineer (Using A.I) a highly robust Regex-based parsing layer that dynamically "carves out" the JSON object from AI text streams.

Managing the React state between our initial "Review" button and the massive 2-column "Review Accordion" caused layout conflicts. I had to strictly separate the component lifecycles and utilize router.refresh() to ensure clean data handoffs (Using A.I).

I initially had the idea of implementing a mermaid graph, this required training models on pre-existing code and diff examples so that each time it "reviewed" work it would accurately map out the workflow of the code and mark errors or bugs. Unfortunately, I was too ambitious and had to scrap the idea to opt for something more functional. I will most likely implement this during a later date.

Accomplishments that I'm proud of

I am incredibly proud of the editor. It's interesting to see it come to life just from a few lines of code, giving a space to edit generated code seems so intuitive and just logical I had to give myself a pat on the back.

The app genuinely looks pretty cool, my only issue is how generic it can come off as.

Training the model to give concise summaries and impact scores that accurately represented the codes influence was something I had a blast doing.

I was very proud of myself for trying to implement multiple models together to have a workflow that would learn from one another to produce an amazing product, this will be a later endeavor.

What we learned

You cannot trust an LLM to follow formatting rules 100% of the time. You always have to second guess and really read what any LLM is pushing out.

The amount of effort that goes into back end and front end design surprised me, I am very excited for whats to come in the future.

Database management and how annoying prisma's can be was an interesting experience (so many hours dealing with failing prismas)

What's next for Revue | GenAI hackathon 2026

Expanding the AI's context window by implementing Retrieval-Augmented Generation across the entire repository, not just the isolated PR diff.

Adding real-time TypeScript/JavaScript color and an split viewer for the code to see side by side implementation of the code.

Bringing the Revue workspace directly into VS Code or any other IDE.

Scaling the program for larger scale use in actual industry environments. E.I lots of model training and big data bases.

Some more quality of life changes to the UX that would really help out.

Built With

Share this project:

Updates

Submission history