Immostory: AI Video Generation That Survives Production
Everyone sells AI video. Almost nobody talks about what happens after the demo: failed renders, retried jobs, duplicated work, and frames that need human eyes. What we learned building a Remotion render pipeline for real estate.
Every AI video demo looks the same: a prompt goes in, a beautiful clip comes out. Then you ship to production and discover the demo hid the entire job. Immostory, a real estate video creation and listing platform, taught us that the render is maybe 20 percent of the work. The other 80 percent is everything wrapped around it.
A completed status is not a delivered video
The first lesson came the hard way: a job marked completed by the render service is not the same as a video a customer would accept. Frames render blank. Transitions stutter on certain property photo aspect ratios. Text overlays overflow on Dutch street names — which are long, and nobody warns you. We stopped trusting provider success responses and built our own verification layer: check the output artifact exists, sample frames for expected content, and only then flip the job to actually-done.
Idempotency or duplicate invoices
Video generation costs real money per render. The moment you have retries — and you will have retries — you need idempotency keys, or you will pay twice for the same video and deliver two slightly different cuts of the same living room. Every job in Immostory carries a stable identity; a retry resumes or replaces that exact job instead of spawning a sibling.
The queue is the product's nervous system
Renders are slow. Customers are impatient. The queue in between has to carry status honestly: queued is not rendering, rendering is not done, and failed-with-retry-in-5-minutes is not failed. We surface the full state machine to the UI because real estate agents plan their listing publishes around it. A vague spinner costs more trust than an honest delay.
What we would tell anyone starting here
Budget the pipeline before the model. The model will get cheaper and better every quarter — your retry logic, your QA sampling, and your job accounting are what compound. That is where the moat is.