🧠 The PDF That Fought Back:

How I Built and Deployed a Narrative Resume Rewriter with GPT, Next.js, and a Very Grumpy TypeScript Compiler

🧠 The PDF That Fought Back:

How I Built and Deployed a Narrative Resume Rewriter with GPT, Next.js, and a Very Grumpy TypeScript Compiler

---

TL;DR:
In between 60-hour workweeks at my full-time job, I built a tool that rewrites resumes into structured 3-act stories using GPT-4. I had no cofounder — just AI and a will to ship. Along the way, I fought PDF parsing chaos, TypeScript errors that refused to die, and UX pivots that made the product 10x better.

---

ACT I — The Idea

This started with a deceptively simple concept:

Upload your resume → get a 3-act narrative built by GPT.
Tailor the tone based on personality and style sliders.
Copy it. Share it. Use it.

I’ve reviewed a lot of resumes in my time, and most of them feel… dry. Even the best ones lack story. So I wanted to give people a way to rewrite their experience as a narrative, not a bullet list.

I had no time. I had no frontend cofounder. I had a demanding job.
But I had GPT… and enough stubbornness to try anyway.

---

ACT II — The Resume Fights Back

The first thing I needed? Text from a PDF resume. That should be easy, right?

Spoiler: it was not.

❌ Attempt 1: pdfjs-dist

• Needed worker threads

• Didn’t work on Vercel

• TypeScript hated it
  → Burned a few hours.

❌ Attempt 2: pdf2json

• Promising at first

• Turned out it couldn’t handle modern PDFs (no real text layer)

• Returned… nothing

✅ Final win: pdf-parse

• Pure Node, no workers

• Just one catch: no type declarations

I had to write a pdf-parse.d.ts file by hand and update my tsconfig just to make the build pass. I also had to do the same for formidable — the popular file upload library — because TypeScript couldn’t make up its mind about default vs named exports.

🧠 Lesson: Don’t fight TypeScript — bribe it with custom types.

---

ACT III — The UI Evolves

At first, the UI was minimal:

• Resume upload

• Textarea

• One slider for GPT temperature

But it felt… flat. It didn’t reflect the personality I wanted the tool to help surface.

So I added:

✅ Personality Sliders

Like “Introverted ↔ Extroverted” and “Practical ↔ Imaginative”
These helped GPT understand who the resume is about.

✅ Narrative Tone Sliders

Like “Structured ↔ Spontaneous” and “Reserved ↔ Expressive”
These controlled how GPT rewrote the resume.

✅ Shareable Copy Button

Instant copy of the narrative for easy pasting anywhere

✅ Dynamic Button Logic

First-time click: “Rewrite It”
Already rewritten? Show “Regenerate”

---

🤖 AI Was My Cofounder

I used GPT-4 not just for the resume rewriting — but for building the tool itself.

It helped me:

• Debug import mismatches between formidable, pdf-parse, and Next.js

• Resolve TypeScript errors about keyof typeof and function typing

• Restructure slider state into generic types

• Draft the backend prompt logic for GPT

Real talk: this tool was built by me and AI — nights and weekends, in between a 60-hour workweek.
No excuses. No full-stack partner. Just a goal and a persistent loop of try → fail → fix → ship.

---

✅ The Final Product

🚀 Live: https://resume-rewriter.vercel.app/
💾 GitHub: https://github.com/KenRambo/resume-rewriter
⚡ Stack:

• Next.js 15 (App Router)

• Tailwind + ShadCN UI

• OpenAI GPT-4 API

• Vercel for deployment

---

⚠️ What Bit Me

• ESLint rules in Next 15 are brutal: no any, no Function, no mercy

• Vercel fails silently on invalid .d.ts typing

• Many PDFs don’t have real text layers (designers, I’m looking at you)

• You need keyof typeof object tricks to use dynamic keys in TypeScript with state

---

🧪 What I’d Do Differently

• Save rewrites to Supabase and generate unique shareable slugs

• Add dark mode and PDF download

• Use @vercel/ai or server actions for GPT integration

• Add GPT autofill for missing resume data

---

✨ What You Can Do Now

• Upload your resume and see it rewritten like a story

• Customize the tone with sliders

• Copy the final narrative with one click

https://resume-rewriter.vercel.app/

If you're building anything with AI — remember:
You don’t need 100 hours. You need a clear idea, some grit, and a robot that writes typesafe code.