How We Built the AI Clip Pipeline Behind ViceForge Clip Machine

Updated October 21, 2026. The first version of this post (May 2026) described an earlier pipeline; the numbers below are current.

The Problem

Creators record hours of gameplay, podcasts and streams, then spend more hours finding a few seconds worth posting. Cutting, captions, reframing and metadata usually live in separate tools. We wanted one pipeline that finds the moments, cuts them and gets them ready to post. That pipeline is now the core of ViceForge Clip Machine.

Five Signals, Not One Score

Listening to speech alone misses most of what makes a gaming clip. Our fusion engine combines five signals:

Vision: the AI looks at a frame every 9 seconds and scores game events such as kills, explosions, chases and clutch plays

Audio: loud spikes like gunfire, explosions and screams

Facecam: we find your facecam and measure how much you react on camera

Speech (Creator plans): the AI reads what is said, in any language, and picks the strongest lines

Moment clustering: a fight or a chase stays one whole moment instead of one loud second

When picture, sound, face and speech agree, the moment ranks higher. Each clip gets a 0 to 100 score with Peak, Action and Speech parts, plus one sentence on why.

Game Profiles

101 ready-made game profiles tune the detection per title. For any other game you type in, the AI writes a profile on the spot and keeps it for everyone after you.

The Stack Today

Next.js in the browser, FastAPI and Celery behind it

AWS Batch with on-demand NVIDIA GPUs for rendering

Amazon Bedrock for vision and text, Amazon Transcribe for speech, Amazon Polly for voice

PostgreSQL and Redis for jobs and metadata

Cloudflare R2 for video storage

Results

$0.14 GPU cost per processed video (a 28-minute input, five 4K clips)

GPU render time down from 26 to 4 minutes, about 8 minutes end to end

Up to 12 GPU jobs in parallel, with no hardware to buy

150+ registered creators and almost 40 paying customers (October 20, 2026)

Key Takeaway

Don't over-engineer the AI. Start with simple signals, measure what real creators keep and post, and add complexity only where it improves the clips. You can try the result yourself at viceforge.app.