Composer by Pixel ML · Built with AWS

The unit of AI video
is not the file.
It’s the cited moment.

A two-hour video is a search problem. Composer finds the exact moment and keeps it tied to the source.
SourceTimestampClaim
Pixel MLComposer by Pixel ML
01 / 07
The problem

Video is everywhere.
Useful video is hard to find.

Teams know the value. They still struggle to pull the exact source moment into the work.
Four evidence signals behind the video activation gap: more than 500 hours of video uploaded to YouTube each minute, 98 percent of enterprise leaders recognize the value of video insights, 45 percent say they have effective extraction tools, and 9 percent say all enterprise data is available and usable for AI. Raw video becomes cited video memory, then Composer turns it into cited, editable work.
Separate studies, not one funnel: YouTube 2021 · Vimeo 2025, n=1,000 · Cloudera 2025, n>1,500.
Pixel MLthe video activation gap
02 / 07
AWS industry use cases

One loop. Four real jobs.

The footage changes. The need to check it does not.
A single Composer video-evidence loop—index, search, cite, produce—supports four illustrative industry workflows: media and entertainment editorial archives, financial services compliance and investigations, retail and CPG store and commerce operations, and travel and transportation safety and field operations.
Illustrative workflows. The stage demo follows the Media & Entertainment path.
Pixel MLone platform · many industry workflows
03 / 07
The agentic workflow

Give it a job. Watch the work.

It plans, calls tools, and leaves an evidence trail.
01 · Goal
Find the moment that makes the point.
02 · Plan
Search, answer, open, collect.
03 · Tools
Search · Ask · Preview
04 · Act
Build a working set, then a draft.
05 · Verify
Open the clip. Editor decides.
plan + tool calls
// Composer MCP — a plan you can inspect
goal: "Find the technical decision"

search_videos({ query: "why did the architecture change?" })
ask({ question: "What tradeoff drove the rewrite?" })
get_preview({ video_id: "current_demo", timestamp_seconds: citation.start })

// Collect moments the editor can check
a real Composer runAABW demo · 01:06
Frame from the recorded AABW Composer demo showing the read-the-video view, a cited source clip, the agent thought-process panel, and the next prompt.
The source, the agent’s work, and the next question stay in one reviewable view.
SourceThought processAsk
Pixel MLGoal · plan · tools · draft · check
04 / 07
Recorded demo
Ask the video.
Check the clip.
Composer finds the exact moment and puts it in the working set.
Open the full run ↗
Pixel MLRecorded product run
05 / 07
Built with AWS
AWS

AWS is doing the work here.

Nova reads the video. AgentCore runs the agent. Every answer points back to the clip.

Public-safe Composer reference architecture: source video is stored in Amazon S3, processed on Amazon EKS, transcribed with Amazon Transcribe, understood with Amazon Nova on Amazon Bedrock, and passed to a Bedrock AgentCore runtime that runs the Composer agent and returns cited moments for human review.
Pixel MLAWS AI/ML in the loop
06 / 07
The ask

Bring one long video.

Leave with the moment you needed and a cut you can keep editing.
Talk to Pixel ML ↗
QR code for the Pixel ML contact page
Scan to talk to us
pixelml.com/contact
Aska real question
Checkthe cited moment
Makethe next cut
Pixel MLComposer by Pixel ML · Built with AWS
07 / 07