From a one-line topic
to a captioned video, automatically
An in-house service that turns a topic into a finished short-form video by running script, voice, video and caption compositing in sequence. It is serverless, and the database pauses when there are no requests, which keeps fixed costs low.
AI pipeline, serverless, video rendering
Real screens.
Captured in September 2026. Select an image to enlarge it.
One short video needs planning, a script, a voice, footage, editing and captions.
A creator who runs a channel alone can hardly repeat that work every day.
Video generation costs money on every call, so the design also had to cut what is spent while nobody is using the service.
We tied the four steps into one pipeline.
A language model writes the script and TTS produces the voice. An image-to-video model generates the footage. Remotion and FFmpeg then add captions and join everything into one video, and that rendering runs on AWS Lambda.
The database is Aurora Serverless v2, which scales to zero when idle. Users top up points and spend them on video generation.
Stack.
Next.js, Hono
The web app and the API are built separately.
OpenAI, ElevenLabs, Kling
Used for script, voice and video generation respectively.
Remotion, FFmpeg
Add captions and join scenes into the final video.
AWS Lambda, SAM
The API and rendering run serverless.
Aurora Serverless v2, Drizzle
A database that scales to zero when there is no usage.
Cognito, Amplify
Cognito handles sign-in and Amplify deploys the web app.