Content Automation Pipeline

What it does

• Generate scene imagesStructured scenes.json drives Gemini / Imagen with a brand-locked flat-vector visual system enforced by an art-director agent’s prompt rules.
• Generate video clipsVeo 3.1 (preview / fast / lite tiers) for motion shots that ride alongside still images.
• Synthesize voiceoverElevenLabs or Gemini controllable TTS, picked at the command line — same scene script, swappable provider.
• Cut captionsOpenAI Whisper into an ASS subtitle file when captions are enabled, otherwise skipped cleanly.
• Assemble the final MP4FFmpeg — image/clip timing, voiceover mix, ambient music bed, optional burned-in captions.
• Tally true costEvery API call’s _cost.jsonl line read through a central pricing table; per-run summary emitted.
• Sync to PostgresOne row per video, one row per API call — for the dashboard layer that watches what each video actually cost to make.

Tech stack

Engineering highlights

• H.01

  Slug-as-unit-of-work

  Every video lives under output/videos/<slug>/ with its own scenes.json, images, clips, voiceover, captions, cost log, and final MP4. A run is fully reproducible from the slug alone.

• H.02

  Provider-pluggable voiceover

  ElevenLabs and Gemini TTS share the same scene contract and output layout; bash run.sh <slug> <voice> [provider] picks at runtime. Adding a third provider is one wrapper module, no core edits.

• H.03

  Cost-as-data, not as logs

  Each generation tool appends a structured JSONL line per API call. A central pricing.py table maps (model, unit, count) → USD. The summarizer reads the ledger; the DB sync replays it into video_costs. Per-run cost is observable, not estimated.

• H.04

  Optional dependencies, hard failures when used

  Whisper, the database, and captions are all opt-in via env flags — but if you opt in and the prerequisite is missing, the run fails loud instead of silently degrading.

• H.05

  Knowledge base as a load-bearing artifact

  raw/ is human source material; wiki/ is the agent-readable distillation. New raw drops are ingested into wiki pages with index.md + log.md updated in lockstep — an LLM-WIKI pattern that keeps the rules current without manual curation.

• H.06

  Agent-authored prep, deterministic execution

  Strategy, Scriptwriter, Art Director, Audio Engineer, and QC agents produce the scenes.json that the deterministic Python pipeline consumes. Creative work is gated and reviewed before any paid API call fires.

What it demonstrates

• Designing a multi-vendor generative-media pipeline (Gemini, ElevenLabs, Veo, Whisper) where every stage is independently runnable, idempotent, and observable — without a job queue, a worker pool, or external orchestration.
• Treating cost as first-class output. Real per-video unit economics fall out of the run rather than being estimated after the fact.
• Separation of concerns between creative spec (agents + wiki) and execution (Python tools + FFmpeg). The script doesn’t know about retention mechanics; the agents don’t know about FFmpeg flags.
• Pragmatic stack choices: shell entrypoint over a workflow engine, JSONL on disk over a queue, Postgres for queryability rather than orchestration, optional Whisper guardrails over heavyweight QA infrastructure.