The Brutal AI: How We Built the Resume Roaster

Job hunting is stressful, and getting generic advice is even worse. Our PM Strategy Department decided to build the AI Resume Roaster to provide brutal, actionable, and Silicon Valley-grade feedback for our users.

The Architecture

We wired up Claude 3 Opus (our preferred model for deep contextual analysis) to a specialized system prompt.

Our prompt engineering went through multiple iterations. Initially, the AI was too polite. We had to enforce a "Gordon Ramsay of Tech" persona. The system prompt forces the LLM to output a precise JSON structure containing:

1. The Roast Score (0-100)
2. Brutal Criticisms (No sugarcoating)
3. Actionable Fixes (How to rewrite the bullet point)

Handling PDF Parsing

One of the biggest hurdles was parsing PDFs cleanly on the server edge. We utilized pdf-parse running inside a serverless function to strip text while maintaining the hierarchy of the resume. If the text extraction fails, our frontend gracefully falls back to a raw text-input mode.

Why brutal?

We learned that users actually prefer harsh feedback when it's accurate. The Resume Roaster doesn't just critique; it rebuilds. It’s an incredibly resource-intensive tool, but seeing users completely transform their CVs makes it worth it.