AD SURGERY - AUTOFIX DEMO

HOW WE AUTOMATED THE EDITOR

1. THE AUTOMATION PIPELINE

We built a multi-agent orchestration pipeline. First, we feed the 41 extracted backend metrics (OpenCV, Librosa) and the original video to Gemini 3.1 Flash Lite. Gemini acts as the "Director," analyzing why the video is a mathematical loser (e.g., 0 hook cuts, too much scene chaos) and writing a precise blueprint for how to fix it.

Once the blueprint is generated, we trigger the Omni Flash API to perform the actual video generation based strictly on Gemini's framing instructions.

2. THE CHALLENGES WE FACED

• AI Hallucinations: Generative video models often hallucinate completely new scenes, losing the original product context or the actor's exact likeness.
• Audio Destruction: Re-generating a video track completely destroys the lip-sync and audio pacing of the original ad, turning high-converting audio into garbage.
• Context Length Limitations: A 30-second ad is often too large to generate in a single reliable shot with high fidelity.

3. THE ENGINEERING SOLUTIONS

• Video Segmentation Protocol: Instead of generating the whole video at once, we use FFmpeg to slice the original video into smaller chunks. We only ask Omni Flash to regenerate the specific chunks that need fixing (e.g., the first 3 seconds to add a hook cut), preserving the pristine original footage everywhere else.
• Audio Ripping & Remuxing: Before touching the video, the pipeline rips the original audio track entirely. After the AI reconstructs the visual chunks and merges them back together, we remux the pristine original audio back onto the final visual track. This perfectly preserves the speech WPM, LUFS, and music, completely bypassing AI audio distortion.
• Structural Storyboarding: We treat the AI output not as a final broadcast-ready ad, but as a "Structural Storyboard." It provides the exact mathematical pacing and cuts needed, acting as a flawless visual blueprint for human editors to easily mimic and deploy.