ThawMail

AI-powered cold emails from any website URL in seconds.

Comment

Inspiration

I've sent hundreds of cold emails as a freelancer. The problem? Generic templates get ignored, but truly personalized emails take 20-30 minutes each to research and write. I'd open a prospect's website, dig through their pages, note their tech stack, find pain points, and then craft something relevant.

I thought: what if AI could do that research and writing in seconds?

What It Does

ThawMail turns any website URL into a personalized cold email. Paste a link, and the AI:

  • Crawls the homepage + 5 key internal pages
  • Captures and analyzes screenshots using GPT-4o vision
  • Identifies the business type, tech stack, and improvement opportunities
  • Extracts contact information
  • Generates a tailored email that references specific details from their actual site

No CRM. No sequences. No complex automation. Just copy the email and send it.

How I Built It

Tech Stack:

  • Next.js 14 + TypeScript frontend
  • Hybrid backend: Express.js for AI operations, Next.js API routes for everything else
  • Puppeteer + Cheerio for website scraping and screenshot capture
  • OpenAI API (GPT-4o-mini for text analysis, GPT-4o for vision)
  • MongoDB Atlas + Upstash Redis for data and caching
  • Clerk for auth, Stripe for payments
  • Deployed on Railway

Architecture Decision: I split AI operations to a separate Express server because Next.js API routes have timeout limits. Website analysis can take 60-90 seconds—Express handles that without breaking.

Challenges

1. Website Scraping Is Hard Every site is different. SPAs, lazy loading, bot detection, cookie banners. I built a multi-layer fallback system: Puppeteer first, then Cheerio, with extended timeouts and retry logic.

2. Keeping Costs Down GPT-4o vision is expensive. I optimized by using GPT-4o-mini for text-only analysis and only calling vision when screenshots add value. This cut costs ~70% while maintaining quality.

3. Credit System Race Conditions Users clicking "generate" multiple times could drain credits incorrectly. Solved with MongoDB transactions and atomic operations—credits only deduct after successful AI completion.

What I Learned

  • Hybrid architectures (Express + Next.js) solve real timeout problems
  • AI vision models are powerful but need cost guardrails
  • The best product isn't the most feature-rich—it's the one that solves one problem really well
  • Cold email personalization at scale is a genuine pain point people will pay for

What's Next

  • Mobile release for Android/iOS

Built With

Share this project:

Updates