Atlas / AI media
An inspectable path from research to video.
Research, narration, media planning and rendering as separate, recoverable stages rather than one opaque generation step.
Public product site
Visit Atlas ↗The problem
Creators working with research-led, narrated video.
A generated video can look convincing while using the wrong visual, an unsupported number or a weak narrative. Fixing one stage should not require treating the whole project as a black box.
How it works
One connected workflow.
- 01Research & writeBuild the narration and segment it into a media plan.
- 02Plan & produceSource or construct visuals and generate clip-level narration.
- 03Review & renderInspect the storyboard, assemble with FFmpeg and optionally publish.
Engineering workflow · a simplified view of the product, not a screen mockup.
Engineering decisions
What matters beneath the interface.
Use a structured media plan between the script and render. Individual shots can be routed to footage, images or constructed graphics; numeric-grounding checks guard chart and statistic inputs.
Research and automated visual checks still need human review. Publishing is an optional, explicit step after rendering, not a claim that AI output is automatically reliable.
My contribution
From product to implementation.
Product architecture, full-stack engineering and the media pipeline across the Angular workspace, Go orchestrator and Python services.
- Versioned declarative visual-rendering contracts
- Per-project script, storyboard and media storage
- Separate narration, rendering and publishing stages
Angular · Go · Python · PostgreSQL · MinIO · FFmpeg