I’ve been working on a small tool called PR AI to help review pull requests faster using AI-assisted context.
The idea is to load the PR diff and let the AI analyze it directly, so you can discuss the changes, risks, and possible improvements without switching between tools.
What it does so far:
Connects to GitHub and loads PR diff context
Import diffs from GitHub, file upload, or paste
Renders readable diffs inside the workspace
Chat with AI using the active PR context
Structured review output (summary, risks, recommendation)
Right now I’m focusing on:
Better traceability from AI comments to exact changed lines
Improving the signal-to-noise ratio in reviews
Keeping the interface lightweight while adding deeper audit signals
It’s still early but already able to load and analyze real PRs.
If you review pull requests often, I’d love to hear:
What would make an AI PR reviewer actually useful for you?
What kind of issues should it prioritize detecting?
I thought about building an PR review powered by AI, and the more I challenged the idea in my mind, the more I think the MoE pattern is the best suited for it to prevent the maximum of hallucinations while creating a virtuous circle of discussions between you and the experts (or even between the experts themselves).
If you go / inspire from this, you may want to start supporting a small subset of technologies like React or Laravel projects, in order to provide the bests set of experts pre-configured with clear prompts so that it gets relevant in PR reviews.
I also think a big part of making this tool useful for entreprise will be the ability to provide some level of customizations, like specific guidelines/linting rules (for example, Domain Driven architecture, or monolithic approaches, etc...