I Used to Manually Scan 300 Resumes. Now I Build AI That Does It Better Than Me

by Pankaj KhuranaApril 30th, 2025
Read on Terminal Reader
Read this story w/o Javascript
tldt arrow

Too Long; Didn't Read

I spent two decades buried in resumes—until I built an AI-powered tool that analyzes resumes with Agentic Vision Models (AVMs). It reads visuals, detects bias, scores candidate fit, and even suggests recruiter prompts.

Companies Mentioned

Mention Thumbnail
Mention Thumbnail

Coin Mentioned

Mention Thumbnail
featured image - I Used to Manually Scan 300 Resumes. Now I Build AI That Does It Better Than Me
Pankaj Khurana HackerNoon profile picture
0-item

I’ve been recruiting for 22 years. Long enough to remember faxed resumes and long enough to say, without irony, “AI saved me.”


In 2010, I’d come home with bloodshot eyes after combing through endless PDFs. One time, I misread a guy’s CV and pitched him for a Python dev role—turns out, he trained dogs. I’ve also seen a candidate include a headshot and a photo of their car. Chaos. And yet, I loved it. Finding the right hire felt like solving a mystery.


Four years ago, I stopped just solving mysteries—I started building the tools.


Today, I lead development at yet to launch product Firki, a system that doesn’t just parse resumes. It understands them. Using Agentic Vision Models (AVMs), Firki acts like an AI recruiter—one that sees messy visuals, extracts real meaning, flags bias, and suggests smart outreach. Think of it as a Chrome extension that whispers, “This resume’s solid, but you might want to ask about their Salesforce experience.”


How AVMs Work: A Recruiter’s X-Ray Glasses

Most resume parsing tech dies when a font gets fancy or someone pastes a logo. Old OCR tools read like they’re drunk—garbling titles, ignoring dates, and vomiting JSON.


AVMs (Agentic Vision Models) change the game. They don’t just scan; they reason.


Here’s how we use them at Firki:



We combine layout-aware vision (like YOLO or LayoutLM) with a language model (GPT-4o or LLaVA) that understands context. The model sees that a name in a fancy top banner is still a name. It detects photos, removes them. Flags bias cues (college logos, addresses, grad years). And if something’s off—like a “Product Owner” label hiding inside a timeline chart—it flags it for review.

The Firki Flow: Real-Time Insights, Not Just Scores

Firki isn’t just parsing. It’s participating.


Here’s what happens when you run our Chrome extension while viewing a job on LinkedIn:

  1. JD gets scraped via DOM parsing (document.querySelectorAll(...))
  2. You upload or auto-import your resume
  3. Firki compares extracted skills, role fit, and seniority using OpenAI or Llama 3
  4. It flags:
    • Missing core skills
    • Visual clutter or bias risks
    • Boolean search terms for recruiters
    • Email prompts to personalize outreach


Example output:

Match Score: 82%

Red Flag:“GTM” keyword missing. Consider adding related experience.

Outreach Prompt: “Hey [Candidate], your Salesforce-heavy background is great, but we’re also looking for someone who’s led GTM rollouts. Mind sharing more on that front?”



Why This Matters: I’ve Been There


I built Firki because I needed it. As VP of Technology at Rocket (a recruiting firm), I saw junior recruiters lose great candidates because:


  • They didn’t understand buzzwords like “GTM” or “ICP”
  • They got distracted by Ivy League logos
  • They skipped candidates who were visually disorganized but actually perfect


AVMs let us focus on the signal, not the noise.


Last year, we processed 10,000+ resumes across three enterprise clients. AVM-led redaction improved bias mitigation by 30%, and our clients saw a 17% improvement in diverse shortlist ratios. One Fortune 500 company called it “a hiring supercharger.”

The Danger: When AI Pretends It’s Perfect

I’ve also seen AVMs blow it.


We once had a resume with a spiral background. Our model misread a “Lead Engineer” as “Intern” because of how the title was rotated. Another time, it flagged a woman’s resume because it thought her portfolio link was spam.


We built in override controls immediately. Every decision is logged, and recruiters can hit “Restore” or “Flag for human review.” AI should assist—not replace—judgment.


Candidates will start asking, “How did the AI read me?” They deserve an answer.

The Future of Recruiting: Craft, Not Chaos

Give it five years, and recruiters will work like product managers:


  • AI scrapes GitHub, Behance, or LinkedIn, highlights standout portfolios
  • You get alerts like:
    “This engineer just shipped a Node.js project with 12K stars. Reach out?”
  • You spend time nurturing, not digging
  • You build hiring strategies, not Boolean strings


Recruiters won’t be résumé scanners. We’ll be storytellers, matchmakers, technologists. AVMs will be the copilot.

Why I Wrote This

Because I’m tired of seeing hiring get framed as “match the keyword” or “beat the ATS.” It should be about understanding people, patterns, and potential. Agentic Vision Models are the first tools I’ve seen that actually think like recruiters. Not perfectly. Not magically. But meaningfully.


At Firki, we’re building this future with real-world hiring at the core. No fluff. No fake-perfect models. Just tools that help you see what matters, faster.


Got a horror story where your parser tanked a great hire? Or building something cooler than AVMs?


Let’s swap notes. I’ve got resumes to stress-test your tech.

Trending Topics

blockchaincryptocurrencyhackernoon-top-storyprogrammingsoftware-developmenttechnologystartuphackernoon-booksBitcoinbooks