ResumeAI

Inspiration

As a 6-year Android developer preparing for my next career move, I spent over 10 hours writing a single resume — and another 2 hours tailoring it for each company. I realized every job seeker faces the same painful process:
manually rewriting resumes, guessing what recruiters want, and never knowing if your resume actually matches the job requirements.

I wanted to build something that eliminates this entirely. What if AI could read your GitHub, understand your skills, and generate a company-optimized resume in 30 seconds?

## What it does

ResumeAI is an AI-powered resume builder that:

1. Drop any link — Paste a GitHub URL and AI instantly analyzes your repositories, tech stack, and project history
2. Company-tailored resume — Paste a job description, and Nova AI generates a resume matched to that specific company's requirements
3. Acceptance roadmap — Get a skill gap analysis, learning paths, and interview preparation guide to actually land the job

The entire flow takes under 30 seconds, replacing 10+ hours of manual work.

## How I built it

### Architecture

• AI Engine: Amazon Nova Lite via Bedrock Converse API with Bearer Token Authentication
• Frontend: Next.js 15 (App Router) + TypeScript + Tailwind CSS
• Auth: Firebase Authentication
• Database: SQLite + Prisma ORM (lightweight, single EC2 deployable)
• i18n: Full Korean + English support with next-intl

### Multi-pass AI Generation Instead of a single AI call, ResumeAI uses a 2-pass generation pipeline:

1. Pass 1: Analyze GitHub profile → extract structured data (skills, projects, experience)
2. Pass 2: Generate company-optimized resume from structured data

This approach significantly improves accuracy and reduces hallucination.

### Anti-Hallucination Validation Nova Lite tends to fabricate metrics (e.g., "10,000 monthly active users") when none exist in the source data. I implemented strict prompt rules:

• COPY EXACTLY rules for names, URLs, and repository data
• Explicit prohibition of fabricated metrics (MAU, downloads, revenue)
• English-only output enforcement to prevent language mixing

## Challenges I faced

### 1. Korean Text Corruption Amazon Nova Lite garbles Korean text when prompts are written in Korean (e.g., "퓨쳐다임" → "이주도"). Solution: Write ALL prompts in English with explicit copy-preservation rules. Korean text in the input is preserved perfectly when the prompt itself is English.

### 2. Hallucinated Metrics Nova would confidently generate fake statistics like "Monthly active users: 1,000" for hobby projects with zero users. Solution: Added NEVER fabricate metrics rules and OVERRIDE ALL PREVIOUS LANGUAGE RULES blocks to suppress this behavior.

### 3. Bedrock API Integration The Bearer Token API uses the Converse endpoint (/model/{modelId}/converse), not the Invoke endpoint. Authentication uses Authorization: Bearer {token} header with 12-hour token expiry. Documentation was sparse, so this required significant trial and error.

What I learned

Prompt engineering for accuracy matters more than model size — Nova Lite with well-crafted prompts outperforms larger models with vague prompts

• Multi-pass generation is essential for structured output — single-pass calls produce inconsistent results
• AI hallucination is a feature-level problem, not just a model problem — you need validation at the application layer
• Bedrock Converse API is powerful but under-documented for Bearer Token auth

## What's next

• LinkedIn and YouTube profile analysis support
• PDF resume export with professional templates
• Real-time job market matching
• Mobile app (Flutter)

Built With

• amazon-bedrock-converse-api
• amazon-nova-lite
• firebase-authentication
• next-intl
• next.js-15
• prisma-orm
• react
• sqlite
• tailwind-css
• typescript
• zod