Parallax

Inspiration

Every builder who has shipped knows this moment: a user drops a furious complaint - "I have no idea what this product even does" - and someone has to manually test the site, write up findings, and file a ticket. Hours of work between feedback and fix.

I wanted to collapse that loop entirely. Build fast, ship often, fix before it hurts.

What It Does

Parallax is an AI agent that takes a raw user complaint, autonomously browses your site through 5 distinct personas, and auto-files a structured UX audit as a GitHub issue - in 60 seconds, zero humans.

Input: "the landing page is confusing, I can't figure out what this does"
Output: GitHub Issue — UX Audit: Score 4.4/10, 5 persona findings, actionable fixes

Also works as a Slack bot - @Parallax https://yoursite.com checkout is broken → analysis posted + GitHub issue filed. Zero clicks.

How I Built It

• Claude Sonnet drives each persona as a browser agent with custom Playwright tools — it navigates, clicks, scrolls, and screenshots like a real user
  
• 5 parallel Playwright instances run simultaneously, each with a distinct persona prompt
• Server-Sent Events (SSE) stream live updates to the frontend as each agent browses
  
• Next.js 16 + React 19 with real-time screenshot display per persona
  
• Slack Events API for the bot — complaint in, full audit + issue link back in-channel
• GitHub API auto-files structured issues with scores, pain points, and screenshots
  
• Vercel AI SDK + Zod for validated structured output across all agents

Challenges

Parallel agent coordination - 5 simultaneous Playwright + Claude sessions. Too many crashed; too few killed the real-time feel. Required careful resource bounding.

Slack's 3-second deadline - forced me to inline the entire analysis pipeline synchronously. No background workers, no queues. Just streaming.

Genuine persona differentiation - getting each agent to truly see the site differently (not just cosmetically varied output) took significant prompt iteration.

What I Learned

• Claude as a browser-use agent (not just chat) is genuinely capable - give it real tools and it navigates with intent
• SSE makes AI feel tangible - watching 5 personas browse live changes how people understand what's happening
• The gap between "AI generates insights" and "AI closes the loop end-to-end" is almost entirely about integrations
  

Mess in. Issues filed. Loop closed. - Built for LovHack Season 2

Built With

• anthropic
• claude
• events
• github
• next.js
• playwright
• react
• sdk
• server-sent
• slack
• vercel
• zod