Prometheus Studio
AI-native video production. Upload. Choose a lane. Ship the next edit.
The Problem
Video editors spend 80% of their time on repetitive cuts, color matching, and caption timing. The creative decisions — the actual art — get buried under technical execution. Prometheus removes the boilerplate so filmmakers can focus on the story.
What I Built
AI Video Pipeline
Transcription (AssemblyAI), content classification, style profiling, caption segmentation, subtitle generation with domain-specific animation (business = pop-up/elastic, lifestyle = fade-in/power2, cinematic = zoom-out/power3).
GPU Render Orchestration
Modal warm pools (keep_warm + memory snapshots), RunPod Serverless for RTX 4090 bursts. Cold start impossible for 2-min render target. H100 for 6DOF matting, A10 for standard motion graphics.
6DOF Video Matting
RECAM Masters 1.2B parameter model for depth-aware subject extraction. Requires persistent warm pool architecture.
Durable Execution Layer
Temporal workflows for crashproof pipeline orchestration. Automatic retry, checkpoint recovery, replaces in-process queue.
Real-Time Collaboration
WebRTC-based review sessions. Upload to R2, stream with inline playback, tap-to-play, double-tap ±10s seek, playback speed control.
Revenue & Outreach System
Hybrid AI/manual lead gen. AI bot handles $997.99 tier volume outreach. Manual surgical outreach (Loom audits, trigger events) for $5K tier. Calendly → Google Meet auto-generated.
The Stack
Impact & Scale
2 min
Target render time for 10-min videos via warm GPU pools.
1.2B
Parameters in 6DOF matting model for depth-aware extraction.
3 Tiers
Revenue modeling across $997, $2,500, and $5,000 packages.
100%
Crashproof durability with Temporal automatic retry & recovery.
Gallery
Lessons & Reflection
- Warm pools are non-negotiable for GPU-heavy AI products. Cold starts kill user trust.
- Temporal durable execution is essential for production pipelines with paying clients. In-process queues die silently.
- The grid UI for structured diversity + human feedback is a teaching tool, not the product. Production is 99% automated with outlier detection.
- Client-specific correction heads (~10MB each) are the real moat, not the base model.