Headlines
  • Now in the lab: Operate, Infra and Encode. Live is next.
  • AI proposes, people approve
  • If you can't rerun it, we haven't finished it
  • Check the grandmaster before you blame the network
  • Be kind, rewind
  • Looking for AV engineers, AI builders and anyone who has kept a plant on air

The AVoAI lab for open media infrastructure.

Mezzframe is an AVoAI lab that researches, builds and documents the use of AI in AVoIP and live media workflows. AVoAI is the AI that works next to AVoIP: agents and models that help engineers build, watch and run their systems. We explore what AI can really do for the technology behind streaming and broadcast, and we share what we learn as open source.

Next

The next shift

Broadcast technology has spent the last 20 years moving from dedicated cabling to software and the cloud, pushed along by streaming. That move is well underway, and the engineering teams at broadcasters are working overtime just to keep these systems running, with little time left to migrate or try new things. We want to help.

Mezzframe wants these teams to have what they need for AVoIP and for bringing AI into their workflows. Once the plant runs as software, AI agents can help run it too: spotting timing drift before it becomes an outage, explaining why a flow dropped, and drafting the change that fixes it. We call that layer AVoAI, and we think it should be designed alongside AVoIP from day one.

Mezzframe's lab runs on AI R&D: agents research, build, review and document. A working engineer directs the agents, runs the operation and publishes the research for the industry to use and learn from. Results ship as open source tools, or as write-ups on what AI can't do yet. We think that matters more than the hype cycle.

  1. SignalAVoIP
  2. MezzanineAVoAI We work here
  3. Peopleengineers, operators
Guide

What we work on

  1. AVoIP

    Next

    Live

    Audio and video over IP for broadcast, streaming and pro AV.

    • ST 2110
    • NDI
    • Dante
    • AES67
    • IPMX
    • PTP
    • SRT
  2. AVoAI

    Now

    Operate

    Agentic workflows for the plant: watch, diagnose, explain, propose.

    • agents
    • telemetry
    • runbooks
    • dry run
    • approvals
  3. Now

    Infra

    The systems and networks media runs on, built with timing and security in mind.

    • automation
    • multicast
    • timing
    • control
    • security
  4. Now

    Encode

    Compression, codecs and color, measured honestly.

    • codecs
    • rate control
    • VMAF
    • learned compression
  5. Later

    Meaning

    Metadata, captions and provenance that travel with the media.

    • captions
    • metadata
    • C2PA
    • understanding
  6. Later

    Preserve

    Legacy video, film and audio, kept safe and restored as labeled copies.

    • restoration
    • archival formats
    • fixity
  7. Later

    Capture

    Cameras, sensors and image processing, as the work gets there.

    • sensors
    • ISP
    • debayer
Rules

How we work

  1. We ship things

    Every project ends in something you can run: a tool, a benchmark, a demo.

  2. Open source by default

    Our public lab code is Apache-2.0 and our writing is CC BY 4.0. A few repos stay private for security, like the one that runs our own lab. If a repo ever needs a different license, it says so up front.

  3. We show our work

    That includes the runs that didn't pan out. If you can't rerun it, we haven't finished it.

  4. AI proposes, people approve

    It's how we run the lab itself. Our agents research, build, review and report. A person reviews and merges every change, and decides anything that touches a live system, a network or an archive.

  5. Built to fit the plant

    We start from open standards and how real facilities behave. Nothing goes in the signal path that we can't explain.

Bench

On the bench

Public repos from the lab. More land here as they're ready.

  • chromaverity

    ◉ Latest

    Check whether a video's color labels match its pixels.

    A command-line check, a CI step, and a read-only MCP server so AI agents can ask too.

    pip install chromaverity
    • broadcast
    • color
    • ffmpeg
    • hdr
    • mcp
    • qc
    • video
    View on GitHub

    PythonApache-2.0Updated

All repos at github.com/Mezzframe

Name

Why “Mezzframe”

A mezzanine file is the high-quality master between the camera original and every delivery copy. It's the middle layer everything else depends on. That's how we approach AI: a reliable, manageable layer between the raw signal and the people producing the work.

Call in

Come say hi

Mezzframe was founded and is run by Mark Jared, a Los Angeles-based media systems engineer and video technologist. We're open to working with engineers and developers across the industry to make AI useful for anyone who has kept a plant on air and the cameras rolling. Industry insight and research shape how we work.

Buy Mezzframe a coffee

Every cup buys us one more ffprobe rabbit hole.