NornPulse

• 

  NornPulse Logo

• 

  NornLabs Website

Inspiration

I’ve always seen short-form video as an amazing gateway. Sure, feeds are full of mindless scrolling, but a great short acts like a spark—pulling someone out of the noise and right down a rabbit hole toward deep, meaningful content like a Carl Sagan documentary. That inspired NornPulse: an autonomous engine built to handle the heavy lifting of finding hooks, vertical cropping, and styling without losing the heart of the original message.

What it does

NornPulse turns long-form videos into vertical shorts, generating each one autonomously and then routing it to a human for approval before anything publishes. Gemini analyzes the transcript and writes the hook, while retention and virality predictions come from real historical benchmarks queried live from ClickHouse—not a model's guess. It automates transcription, cropping, and dynamic kinetic subtitles, all inside a smooth Streamlit dashboard complete with human-in-the-loop review and real-time ClickHouse analytics.

How we built it

We designed it around a trio of specialized agents—Urðr for analytics, Verðandi for orchestration, and Skuld for rendering. Gemini Flash handles the intelligence and transcript analysis, a custom Python and FFmpeg pipeline takes care of the video processing, and Streamlit ties it all together into a clean, responsive interface.

Challenges we ran into

Timing subtitles was a massive headache. Cutting a clip at a specific timestamp shifted the video timeline to zero, but our subtitle files still held absolute timestamps, causing captions to completely vanish. We had to write a custom re-basing script to shift and clamp those windows correctly. Plus, our rapid UI testing accidentally torched our API rate limits until we set up local file caching.

Accomplishments that we're proud of

We actually pulled off a seamless blend of heavy video engineering and generative AI, building a tool that goes from raw long-form video to a polished, subtitle-ready vertical short without a single manual edit.

What we learned

AI is a brilliant assistant, but it shines brightest with human guidance. By adding a quick training feedback loop to our review queue, NornPulse doesn't just cut clips—it learns from every rejection to get smarter over time.

What's next for NornPulse

Next up, we want to expand multi-platform auto-publishing, open up support for more source types, and tie the feedback logs directly into automated agentic memory updates.

Built With

• clickhouse
• cloud-run
• ffmpeg
• google-cloud
• google-gemini
• plotly
• python
• streamlit
• vertex-ai
• youtube