Case study

Postcrest: building an AI content and scheduling platform

Postcrest is the AI workspace for social media creators we've been building since 2023. It generates photos, videos, voice-overs, music and captions, then schedules and publishes them to Instagram, YouTube and X, with comments and insights in the same place. This is how it's put together, and what we learned along the way.

Product
AI content and scheduling SaaS
Building since
2023
Stack
Rails 8, Hotwire, PostgreSQL
Languages
English and German
Status
Live, paid plans only
Postcrest's multi-shot video planner, where AI writes a description and a prompt for every shot before anything is rendered
The challenge

The challenge: ten tools for one post

Postcrest started with a frustration Martin had as a creator: "I was paying for ten different tools, copy-pasting between five tabs, and still not posting consistently." One app for images, another for video, another for voice-overs, another for scheduling, and none of them talked to each other.

The goal was one workspace where content is created and published in the same flow. That meant integrating a fast-moving set of AI providers, the strict publishing APIs of the big networks and a billing model fair enough to cover very different costs per generation.

Our approach

Our approach

We built Postcrest as one Rails application with a server-rendered Hotwire interface, background jobs for everything slow and a credit system between the user and every model. Features shipped in small steps: Stripe billing in early 2024, TikTok in 2024, background jobs on Solid Queue in 2025, publishing to X and YouTube in 2026.

  • One credit balance across every AI model, with the cost shown before each run
  • Every provider behind its own generation class, so models can be added or retired without touching the product
  • Publishing through each network's official API
  • Server-rendered pages with live updates through Turbo Streams, without a separate front-end app

20+ models behind one interface

Postcrest offers more than 20 models for images, video, music and talking avatars, from providers including OpenAI, Google, ElevenLabs, Runway, MiniMax and Kling, reached through their own APIs and through Fal and Replicate. Users pick a model per task in one picker that shows the credit cost of every run.

Behind the picker sits a model lineup registry that describes each model: what it does, what it costs and which inputs it needs. The marketing pages read their model counts from the same registry. It's the same approach we use when we build generative AI apps for clients: the product stays stable while the models change underneath.

Postcrest's video model picker listing Grok Imagine, Gemini Omni Flash, WAN 3.0, Kling 3.0 and MiniMax with their credit costs
The video model picker: one list across providers, each model with its cost in credits.

Credits that stay fair

A caption, an image and a ten-second video cost wildly different amounts to generate. Postcrest prices every run in credits, calculated from rules per model: resolution multipliers, a rate per second of video and add-ons. The interface shows the total before the user clicks, and a failed generation costs nothing.

Credits come from the subscription, which resets every month, and from top-up packs that expire after a few weeks. A ledger records every credit by source and expiry date, which keeps refunds, disputes and support questions simple. We wrote up the pattern in our guide to usage-based billing with Stripe credits.

Jobs that hang for an hour are failed and refunded automatically, and a refusal by a content filter is retried once on a more lenient model before the user sees an error.

Postcrest's pricing breakdown before generating a video: a base cost plus a rate per second for a five-second 1080p clip
The cost breakdown shown before a video is generated.

Publishing to Instagram, YouTube and X

Every network has its own rules, so posts are checked per platform before they're queued: caption and hashtag limits and carousels of up to ten items on Instagram, Reels re-encoded with ffmpeg to stay within Instagram's bitrate limit, resumable uploads with native scheduling on YouTube, threads and chunked media uploads on X. A queue worker looks for posts to publish every 15 seconds.

TikTok publishing is built on the Content Posting API. Until TikTok's app audit passes, posts are forced to private drafts and new connections stay hidden, so the product never promises more than the platform allows.

Tokens are refreshed per provider before they expire, under a row lock so two jobs never refresh the same token at once, and an alert fires when the share of successful connections for a network drops. Instagram comments sync into Postcrest, where they can be answered, hidden or deleted, with AI-drafted replies the user approves first. We bring the same groundwork to every social media platform we build.

Postcrest's posts overview: scheduled Instagram carousels with their captions, publish times and status
The post queue, with scheduled Instagram carousels (test account name blurred).

Workflows that chain models

Some content needs several models in a row. Multi-shot video plans every shot with AI before rendering. Consistent characters are trained once and reused across photoshoots and videos, and voices can be designed from a description or cloned, then attached to a character. Captions, text overlays and subtle zooms are rendered with ffmpeg.

Every step runs as a background job with live progress in the interface, so a long workflow never blocks the page and a failed step can be retried on its own.

Postcrest's Design Your Voice screen with a written voice description and a progress bar through the voice settings
Designing a voice from a written description, one step in a character workflow.

Built to be found

Postcrest is built to be found through programmatic pages: 98 free browser tools in 12 categories, 39 comparison and 39 alternative pages, use-case pages per network, plus an academy and a blog. Sitemaps are split into families, an llms.txt describes the product for AI assistants and IndexNow announces new pages.

Page titles are tested against Search Console data, and product experiments divide traffic with a Thompson-sampling bandit, which shifts visitors towards the variant that performs best. The same thinking goes into the multilingual and programmatic SEO we build for clients.

Results

Postcrest in numbers

commits since 2023
1,700+
AI models in one workspace
20+
free browser tools as landing pages
98
automated tests
1,400+

Stack

One Rails application rendered on the server, with background jobs in their own container and deploys with Kamal on European servers behind Cloudflare.

  • Ruby on Rails 8
  • Hotwire
  • ViewComponent
  • Tailwind CSS
  • PostgreSQL
  • MongoDB
  • Solid Queue
  • Stripe
  • ffmpeg
  • Kamal
  • Hetzner
  • Cloudflare
What we learned

What we learned

Platform reviews are part of the timeline

Publishing to TikTok is built, but TikTok's app audit decides when it can go public. Until it passes, posts go out as private drafts. That's why we plan the reviews by Meta, Google, X and TikTok into every social media project from day one.

Fewer moving parts

When Postcrest moved to Kamal, Redis went with it: background jobs, caching and websockets now run on Solid Queue, Solid Cache and Solid Cable on top of PostgreSQL. One service less to run, monitor and back up.

Never hard-code what changes every month

AI models arrive and retire every few months. Model counts on the marketing pages come from the same registry as the model picker, and a test fails if anyone types a fixed number into the copy. The site can't promise a model the product no longer offers.

Building a content or creator platform?

Tell us what your users should create and where it should go. You'll get an honest view on models, APIs and what to build first.

martin@balk.dev

Typically a reply within one business day.

Useful to mention

  • What you want to build, in a few sentences
  • Who is going to use it
  • Any deadline or launch date
  • Tools and systems it should connect to