Back to blog
EngineeringFebruary 21, 202612 min read

How We Built a Full SaaS in 24 Hours with Claude, Next.js, and Supabase

From concept to live product — with AI debriefs, audio generation, payments, podcast search across 4.5M+ shows, and 12 pages — in a single day. Here's exactly how.

The Challenge

We wanted to test whether AI-assisted development could take a product from idea to live, paying-ready SaaS in 24 hours. Not a landing page. Not a prototype. A real product with:

  • • Full authentication (email, Google OAuth, magic links)
  • • AI-powered content pipeline (transcription → analysis → audio)
  • • Payment integration with real payment processor
  • • Interactive audio player with highlights, chapters, search
  • • 4.5M+ podcast catalog with real-time search
  • • Email system, admin dashboard, legal pages, blog
  • • Custom domain, CDN, and email delivery

The Timeline

10:00

Concept & Architecture

Discussed podcast summarizer ideas, researched competitors (Snipd, Podwise, Castmagic), mapped out the stack. Decided: Next.js 16 + Supabase + Cloudflare R2 + Vercel.

11:00

Database & Backend

Designed 8 Supabase tables with RLS policies, auth triggers, and tier enforcement. Built the worker pipeline: PodcastIndex → YouTube/Whisper transcription → Claude analysis → TTS audio → R2 storage.

14:00

First Debrief Generated

End-to-end pipeline test: All-In Podcast E261 → 100MB audio → 8 Whisper chunks → 71K char transcript → Claude debrief with 7 takeaways and 6 quotes → TTS audio on R2. Cost: $0.65.

16:00

Frontend MVP

Landing page, auth (email + Google OAuth + magic link), dashboard with podcast search, debrief viewer, pricing page. All in Next.js 16 with Tailwind CSS v4.

19:00

Payments & Polish

Stripe payment integration, admin dashboard, explore page (4.5M+ podcasts), episode intelligence (interactive highlights, chapters, keyword search), about page.

22:00

Domain, Email & Deploy

Bought poddebrief.com, configured DNS on Cloudflare, set up Resend for transactional emails, deployed to Vercel, Supabase SMTP configured, custom auth emails.

02:00

Audio CDN, Podcast Show & Blog

Cloudflare Worker for R2 audio delivery, generated PodDebrief Weekly Episode 1 (~20 min), RSS feed, legal pages, blog. Switched TTS from ElevenLabs to OpenAI (90% savings).

The Stack

ServiceRoleCost/ep
Claude Sonnet 4.6AI debrief generation + episode intelligence~$0.22/ep
Next.js 16Full-stack frontend (App Router, React Server Components)
SupabaseDatabase, auth, RLS, real-timeFree tier
VercelFrontend hosting, serverless API routesFree tier
Cloudflare R2 + WorkersAudio storage + CDNFree tier
OpenAI WhisperAudio transcription (when YouTube unavailable)$0.10/ep
OpenAI TTS (tts-1)Audio debrief generation$0.029/ep
PodcastIndex APIPodcast catalog (4.5M+ shows)Free
StripePayments (Cards, Apple Pay, Google Pay, iDEAL, SEPA)Transaction fees
ResendTransactional email (magic links, digests)Free tier

Total cost per debrief: ~€0.35. Fixed monthly costs: ~€43 (at scale with Vercel Pro). Most services on free tier during launch.

The AI Pipeline

The core magic is the debrief pipeline. When a new episode drops on any podcast you subscribe to, here's what happens:

1.

Detect — PodcastIndex API alerts us to new episodes

2.

Transcribe — Try YouTube transcript (free, 80% hit rate) → fall back to OpenAI Whisper with ffmpeg chunking for large files

3.

Analyze — Claude Sonnet 4.6 generates structured debrief: takeaways, quotes, predictions, action items, links

4.

Intelligence — Second Claude pass extracts highlights (8 categories), chapters, keywords

5.

Voice — OpenAI TTS converts the debrief to audio (auto-chunked at 4096 chars)

6.

Deliver — Audio to Cloudflare R2, debrief to Supabase, notification to user via email + RSS

Lessons Learned

1. AI doesn't replace architecture decisions

Claude wrote most of the code, but the hard decisions — Supabase vs. Postgres, R2 vs. S3, Supabase vs. Firebase, volume-only pricing — still required human judgment. AI accelerates execution, not strategy.

2. YouTube transcripts saved us 80% on Whisper costs

Most popular podcasts publish on YouTube. We always try the free YouTube transcript first. Only ~20% of episodes need Whisper transcription.

3. OpenAI TTS is 90% cheaper than ElevenLabs for this use case

We started with ElevenLabs ($0.30/debrief) and switched to OpenAI tts-1 ($0.029/debrief). For informational audio, quality difference is negligible.

4. Vercel serverless can't do everything

We hit issues with aws4fetch for R2 signed requests on Vercel's serverless runtime. Solution: Cloudflare Worker with R2 binding — faster, simpler, and free tier covers our usage.

5. "Same features. Just more." is the simplest pricing strategy

Every tier gets every feature. The only difference is volume (5/50/200 debriefs per month). No feature gates, no confusion, no "but I can't do X on the free plan" complaints.

What's Next

We shipped the product. Now comes the hard part: getting users. Our roadmap includes PodDebrief Weekly on all podcast directories, a weekly newsletter, Notion export, and eventually a mobile app. But the core product — AI debriefs with interactive highlights — is live and free to try.


Try PodDebrief Free

The product we built in 24 hours. 5 free debriefs per month. All features. No credit card.

Start for free →