DiffWatch

Overview

Ingests a pull request and returns actionable feedback in under 15 seconds, running parallel analysis across diffs, CI status, vulnerability signals and bundle impact, then surfacing it through four distinct reviewer personas.

The problem

Code review is one of the highest-leverage activities on a team, but it's slow and inconsistent. Reviewers focus on different things, miss security issues, or skip context entirely. I wanted the same thorough multi-angle review to happen automatically in the time it takes to open a PR.

Approach

A fan-out pipeline in Next.js runs diff parsing, CI API calls and vulnerability pattern matching simultaneously. Each prompt is scoped to a specific reviewer persona so the feedback has a distinct, purposeful voice instead of arriving as one blob of AI text. Streaming over Server-Sent Events keeps the UI responsive while analysis runs.

What it does

• Parallel analysis pipeline across diffs, CI status, vulnerabilities and bundle impact
• Four scoped reviewer personas — Architect, Security, PM, Intern — with streaming output
• Commit-history analysis, risk detection and anti-pattern flagging

What I learned

Prompt scoping is everything. Giving the model a narrow, well-defined role — security reviewer versus PM — produces far more useful output than asking one general question. The parallel fan-out also cut latency roughly 60% against sequential calls.