AI experiments
AI experiments
Sometimes I deliberately build something "playful" to demonstrate what's possible when you automate a process. For example: converting a blog post into a short video with scenes, images, and voice.
This is only useful when you can ensure quality. That's why I build these things with measurable checks and clear boundaries.
Example: blog → video
What happens:
- Blog text is split into scenes (short segments with one message).
- For each scene, an image prompt is created and an image is generated.
- For each scene, text-to-speech is generated. The voice in these videos is an AI-generated version of my own voice.
- Everything is assembled into one video.
Why this is interesting:
- a long text becomes faster to consume
- you can reuse content in multiple forms
- you immediately show that automation isn't just "code", but also presentation
What to watch out for:
- voice can sound "scripted" → adjust pacing/intonation
- images can distract → choose calm and consistency
- content must be accurate → text remains authoritative, no hallucinations
A real run, with numbers
No mockup — these are the actual numbers from one run of this pipeline on a blog post (privacy-first AI for sensitive data, with the video at the bottom):
- 8 scenes from one source post
- 8 AI images (Gemini), brand-consistent via a fixed style contract
- ~1,250 characters of text-to-speech in my cloned voice
- 1 script call (gpt-5.2) that splits the text into scenes
- 3 languages (NL/EN/DE) from the same source; language-independent images are shared via caching
- a ~1.5 minute video, "playful-office" style
Usage is measured per run, so cost scales with your own rates instead of being a black box. That's the difference between a fun demo and something you put into production under control.
Want to know if this could be useful for your content or process? Put an example (text or workflow) to Elk Solutions. The first substantive response comes from the people who built it.