Enhance Resume Parsing with AI and Gemini 2.5 Flash

[MoonLawlier]

Pull Request: AI-Assisted Resume Parsing with Gemini 2.5 Flash

Description

Fixes #7984

Implemented an AI-assisted resume parsing pipeline using Gemini 2.5 Flash to significantly improve profile extraction accuracy across diverse resume formats while maintaining a lightweight local-first architecture.

What was added

• Integrated Gemini 2.5 Flash for structured resume data extraction.
• Added JSON-based profile parsing workflow.
• Implemented schema validation for AI responses using Zod.
• Added automatic profile form population from parsed data.
• Added localStorage persistence for parsed profile information.
• Implemented fallback mechanism to the existing rule-based parser when AI parsing fails.
• Improved support for resumes created with Canva, Figma, Microsoft Word templates, and other modern resume builders.
• Prevented raw PDF metadata and object references from being incorrectly displayed as profile information.

Why this change?

The existing PDF text extraction approach struggles with many real-world resume formats, often resulting in:

• Missing names and contact information
• Empty education and experience sections
• Incorrect skill extraction
• PDF object references being interpreted as user data
• Poor compatibility with visually designed resumes

This enhancement improves extraction accuracy while preserving a lightweight architecture and avoiding unnecessary backend complexity.

---

Pillar

• 🎨 Pillar 1 — New Theme Design
• 📐 Pillar 2 — Geometric SVG Improvement
• 🕐 Pillar 3 — Timezone Logic Optimization
• 🛠️ Other (Bug fix, refactoring, docs)

---

Visual Preview

Previous Output

Full Name: N/A

Skills:
Parent 7 0 R
Prev 13 0 R
...
endobj
15 0 obj

New Output

{
  "name": "John Doe",
  "email": "john@example.com",
  "skills": ["React", "TypeScript", "Next.js"],
  "education": [
    {
      "institution": "ABC University",
      "degree": "B.Tech CSE"
    }
  ],
  "experience": [
    {
      "company": "XYZ Ltd",
      "role": "Frontend Developer"
    }
  ]
}

Parsing Flow

Resume Upload
      ↓
PDF Text Extraction
      ↓
Gemini 2.5 Flash
      ↓
Schema Validation (Zod)
      ↓
Profile Generation
      ↓
localStorage Persistence
      ↓
Auto-filled Profile Form

Fallback Flow

PDF Text Extraction
        ↓
Gemini Parsing
        ↓
Success → Structured Profile

OR

Gemini Failure
        ↓
Existing Rule-Based Parser
        ↓
Best-Effort Profile Generation

---

How I solved this issue:

• Added AI-powered resume parsing using Gemini 2.5 Flash.
• Implemented structured JSON extraction and validation.
• Added profile persistence using localStorage.
• Added graceful fallback to the existing parser to ensure uninterrupted user experience.
• Improved compatibility with modern resume templates and complex layouts.

[vercel]

@MoonLawlier is attempting to deploy a commit to the jhasourav07's projects Team on Vercel.

A member of the Team first needs to authorize it.

[retenta-bot]

This PR significantly enhances the resume parsing capabilities by integrating Gemini 2.5 Flash, which should address many of the issues we've faced with the existing parser, particularly with complex resume formats. It also includes a fallback mechanism to ensure reliability, similar to past decisions we've made regarding garbage character handling in the parser. Looking forward to your feedback on this improvement! 😊

[Aamod007]

Awesome integration of Gemini 2.5 Flash! The addition of the AI fallback mechanism in
esume-parser.ts\ makes the parser much more resilient to differently formatted resumes.

I particularly like how you robustly handled extracting and validating the JSON response from the LLM using \zod\ and checking multiple possible payload structures (\ ryParseCandidate). The \sanitizeExtractedText\ function is also a smart addition to clean up PDF stream artifacts before feeding the text to the model.

Labels applied:

• type:feature: Introduces AI-assisted resume parsing.
• level:advanced: Implements LLM API interactions, robust schema validation, and fallback logic.
• quality:clean: Well-structured code with good test coverage for the PDF stream issue.

[github-actions]

⚠️ Hey @MoonLawlier, this pull request has been inactive for 3 days. It will be automatically closed in 2 days if no further activity occurs.

If you are still working on this, please push your latest changes or leave a comment to keep it active.

[MoonLawlier]

hi @Aamod007 I didn't really understood what exactly I need to improve or change in my current code?