Case study / Ellie
Ellie, the AI agent that watches video so editors do not have to.
Long-form video in, ranked clips out. Ellie analyzes scenes, transcribes speech, scores moments against editorial rules, and ships the best segments to a production pipeline.
02 / Problem
Editors spend hours scrubbing for ten seconds.
Podcast clips, sales-call highlights, security footage review. The valuable seconds hide inside hours of footage. Manual review does not scale, and most 'AI video' tools stop at autoscroll.
03 / Approach
Scene-aware analysis with a scoring rubric.
Scene detection
Visual and audio change-point detection to chunk the timeline.
Transcription
Open-weight ASR with speaker diarization for cost-efficient throughput.
Editorial scoring
Claude scores each chunk against a domain-specific rubric. Highlights surface ranked, not random.
Clip export
Ranked clips written back into the customer pipeline, ready for human review or auto-publish.
04 / Outcome
Hours of footage to a ranked shortlist in minutes.
Used in podcast post-production and internal review workflows. Scales horizontally for batch jobs.
FAQ
Common questions.
- What does Ellie do?
- Analyzes long-form video at scale. Detects scenes, extracts spoken content, scores moments against editorial rules, and outputs ranked clips.
- What model powers it?
- Claude for reasoning and scoring, plus open-weight transcription models for speed. The architecture is model-agnostic.
- Can I use it on my footage?
- Engagement-based. The audit is the way in.
Have video your team cannot keep up with?
Scoping the data, the rubric, and the pipeline runs through Systems Decision Audit at TechTide AI. Technical write-ups live on alexcinovoj.dev.