Production Pilot

• 

  Production Pilot

Inspiration

Production Pilot came from a real production problem: self-recorded videos often arrive with avoidable issues such as wrong duration, poor framing, bad lighting, or unusable audio. Those problems create another round of feedback and recording for both the performer and producer. We wanted creators to catch them before submitting.

What it does

Production Pilot reviews a self-recorded video against the requirements of a specific production. It returns either READY TO SUBMIT or NEEDS CHANGES, explains exactly what to fix in the next recording, and can compare multiple takes to recommend the strongest passing one.

How we built it

We separate objective measurements from visual judgment. ffmpeg and ffprobe check things like duration, format, and audio. Six sampled frames are analyzed by GLM-5.3-Flash on Nebius Token Factory, then NVIDIA Nemotron-3-Super receives those observations, the measurements, and the production requirements to make the framing and lighting decisions and write corrective instructions.

Challenges we ran into

The hardest part was reliable visual judgment. The NVIDIA models available to us on Token Factory did not accept image input, so we built a two-stage architecture instead. We also found that vision models could misread framing when given several frames at once, so we switched to focused, per-frame observations before letting Nemotron make the final decision.

Accomplishments that we're proud of

We built a working end-to-end reviewer rather than a scripted demo. The three demo takes repeatedly produce the expected results, and four separate real submissions also matched both the expected verdict and primary failure reason in the final regression runs. The project also includes two different production profiles, a traceable review pipeline, and automated tests covering measurements, verdicts, model-output handling, streaming, and cleanup.

What we learned

The biggest lesson was that not every production rule should be handled by AI. Duration and format are better measured directly, while framing and lighting need visual judgment. Separating those responsibilities made the system more reliable, easier to explain, and easier to adapt to different production briefs.

What's next for Production Pilot

Next, we want to test Production Pilot with more real productions and production profiles, including self-tape auditions, remote conference recordings, creator campaigns, and other workflows that depend on self-recorded video. We also want to deploy it as a secure hosted service so creators can check a take from their phone before sending it to a producer.

Built With

• creatortools
• nebius
• nemotron
• productionai
• remoteproduction
• selftape
• videoqa
• videoreview