Prototype · Independent · Applied AI · Sep 2025 — present
AI video workflow
A FastAPI control plane that plans and assembles short concept ads, with paid model calls locked until a person has reviewed the dry-run.
Workflow architecture · Illustrated from the case study.
Result
A FastAPI worker exposes dry-run, scene generation, job status, review, technical QA, and FFmpeg assembly endpoints. Pydantic and JSON Schema constrain inputs; configuration records claims, assets, hashes, costs, and delivery profiles.
Problem
Generation is expensive and marketing copy lies easily. A pipeline that always calls a model is a spend leak. A pipeline that cannot show why a line is forbidden is not reviewable.
Role
Independent. I designed the worker control plane, review boundary, configuration model, FFmpeg assembly, and run metadata. A person reviews the dry-run before approving paid generation.
Decisions
Dry-run is the default. Forbidden claims block approval. Scene generation is reviewable one shot at a time, and an optional inactive n8n template does not own the core workflow. The current delivery profile prefers 18-second concepts, with a 15-second fallback.
Evidence
Claim lint is a public rule-check demo with no API key. The private project contains the worker control plane and locally generated concept outputs; this page currently shows an architecture illustration rather than a live generator or concept film.
Limits
Prototype at a frozen P0 stage, not a public generation service or commercial campaign system. The optional n8n templates are inactive, and this case does not claim a completed dubbing or lip-sync pipeline.