PostyPop
One piece of content → 22 platform-tuned variations in seconds. Live SaaS, real customers.
The AI part
Platform-specific prompt engineering (character limits, hashtag density, voice) layered on a four-stage fallback chain — Groq Scout → Llama 3.3 70B → Compound → GPT-4o-mini — so output stays consistent regardless of which model executes the request.
Stack
Why I built this
Creators rewrite the same post for LinkedIn, Twitter, Instagram, TikTok, Threads, Bluesky, Pinterest, Mastodon, Discord, YouTube, and Facebook — by hand. Eleven platforms, eleven voices, eleven character limits. The work is real and the LLMs are good enough; nobody had stitched it into a clean product yet.
How it works
- Single input, 22 outputs. Two A/B variations per platform, each tuned to that platform’s conventions (LinkedIn formal, Twitter punchy, Instagram hashtag-dense, TikTok hook-driven).
- Multi-model fallback. Groq’s Llama 4 Scout for speed; if quota hits or output fails validation, automatic failover through Llama 3.3 70B → Compound → GPT-4o-mini. The user never sees a “rate-limited” error.
- Brand voice. Higher tiers learn a tone from samples and apply it across every variation.
- Chrome extension. Highlight text on any page, generate, copy, paste. Manifest V3, approved on the Chrome Web Store.
- End-to-end SaaS. Sign-in (Google + Apple OAuth, email/password with bcrypt), email verification, Lemon Squeezy webhooks for the full subscription lifecycle, audit log, GDPR export.
What it took
- A 741-line CLAUDE.md of operational lessons — including the security work: SSRF guards on user-supplied URLs, HMAC webhook verification, constant-time password comparison, parameterized SQL.
- Production deployment on Hostinger Business with a
start.mjsbootstrap. - llms.txt + JSON-LD + OG + sitemap, because launching a SaaS in 2026 means launching for AI crawlers too.
- Splash animation built with GSAP (logo draws, fills, pops, brand reveal) — a one-time joke that took longer than it should have, and earned its keep.
What I learned
Shipping a SaaS in public teaches you that the LLM is the easy part. The actual work is OAuth flows, webhook idempotency, abuse prevention, and the unglamorous middle layer that keeps a paid customer from rage-quitting on a Sunday morning.