A one-command, human-supervised AI pipeline that produces hour-long lofi YouTube videos end to end — artwork, motion, music, render, metadata and upload — governed by a strict $2.00-per-run cost ceiling, with a zero-dependency Studio dashboard for full observability.
Run python orchestrator.py --theme "rainy night" --season autumn and a complete lofi video is generated, rendered and (optionally) published — all under human supervision and a hard spend ceiling. Its mascot is Porsuk, a headphone-wearing night-owl cat.
Image+motion and music generation run concurrently, then merge, metadata and upload. Coordinated by a single orchestrator with checkpointed state.
A premium, zero-dependency (stdlib only) observability panel: sessions, cost analytics, live pipeline state, artwork gallery and logs.
Every run is bounded by a $2.00 ceiling. Spend is tracked per asset (image / video / music) and the pipeline halts before it can exceed the cap.
fal.ai MiniMax Music v2 for instrumental lofi tracks, AI image + motion generation for the Porsuk cover loop. All curated by a human.
Uploads default to private. Secrets are git-ignored. The dashboard redacts private local paths and never reads credential files.
Each step is saved to state.json. Interrupted runs continue where they left off, and completed steps are never re-run.
Cinematic dark interface, glassmorphism panels, animated metric cards and premium charts — running locally on the Python standard library, with no external dependencies.
Overview tab — hero, animated KPIs, cost-per-run chart, upload-success ring and seasonal distribution. Real data from the deployed pipeline.
Every run advances through a 7-step state machine. Click a step to see what happens at each stage — the same rail is animated live inside the dashboard.
Phase 1 (image+motion ‖ music) runs in parallel, then merge, metadata and upload run in sequence.
Each run is capped at $2.00. Spend is recorded per asset and surfaced live. Drag the slider to see how the governance bar reacts as a run accumulates cost.
The dashboard reads only local session files and logs — nothing is generated, uploaded or deleted. Switch between the surfaces below.
Animated metrics, cost-per-run chart and upload-success ring.
A small, type-aware Python codebase: one orchestrator, five specialised agents, shared utilities, and a stdlib-only dashboard.
Real, unedited captures from the running dashboard on actual session data — click any to enlarge.
Filterable table · status badges · CSV/JSON export
Animated 7-step rail · cost-cap bar · activity log
Porsuk artwork · YouTube / loop / 1h ribbons
YouTube embed · artwork · tracks · timeline
Level badges · session markers · redacted paths
Launcher modal · theme / season / atmosphere · $2.00 cap up front