📝 Salvo

Brief in → scored ad variants out — best three ship

Brief in → N scored ad variants out → best three ship. Salvo is a batch creative factory built on Genblaze + Backblaze B2: give it a creative brief, it fans out N ad-image variants through one real Genblaze Pipeline, stores every variant (with provenance) through a B2 object sink, scores them with an explainable, deterministic ranking, and surfaces the top 3 to ship.

⚙️ How it works

flowchart TD
    A([Brief]) --> B["N × MockAdProvider steps<br/>one PNG + headline each"]
    B -->|"genblaze Pipeline.astream(max_concurrency=N)<br/><b>real fan-out</b>"| C["ObjectStorageSink<br/>LocalDirBackend · HIERARCHICAL<br/>campaigns/{date}/{run}/…"]
    C -->|"read_manifest(verify=True)<br/><b>provenance read back</b>"| D["ranking.rank_variants(brief, variants)<br/><b>explainable scores</b>"]
    D --> E([Top-3 — the three highest-scoring variants])

Loading

Every variant is scored on three transparent signals, and each score ships with a plain-English breakdown of exactly how it was reached:

Signal	Weight	What it measures
Brief coverage	0.45	how many of the brief's keywords the headline actually uses
Headline length	0.25	closeness to a 22–42 character scannable sweet spot
Engagement index	0.30	a deterministic pseudo-signal seeded from the content hash

Honesty note. The engagement index is not real click/CTR data — it is a deterministic stand-in seeded from the content hash so the ranking is reproducible offline, and every reason string says so. No fabricated metrics are presented as real.

🟢 OFFLINE mode (the always-green demo path)

Salvo runs with zero credentials by default. OFFLINE=1 (the default) uses a mock image provider that emits real PNG bytes (via a dependency-free raw-PNG encoder — no Pillow, no ffmpeg) and an on-disk LocalDirBackend that implements Genblaze's documented StorageBackend interface. The full pipeline — fan-out, storage, manifest verification, ranking — exercises real Genblaze code paths without touching the network.

Setting B2_KEY_ID / B2_APP_KEY switches storage to a real Backblaze B2 bucket via Genblaze's S3StorageBackend; the app auto-detects credentials at startup.

🚀 Quickstart

uv sync --extra dev
OFFLINE=1 .venv/bin/python -m pytest  # 37 tests, all green
OFFLINE=1 .venv/bin/python -m uvicorn app.main:app --port 8000

Then open http://localhost:8000/console — type a brief, hit Generate, and watch the scored variant grid render with the top 3 highlighted.

API

Method	Path	Purpose
GET	/healthz	liveness + mode + genblaze version
POST	/campaigns	{ "brief": "...", "n": 6 } → variants + ranking + top-3
GET	/campaigns/{id}	fetch a completed campaign
GET	/campaigns/{id}/variants/{i}.png	a variant's PNG bytes
GET	/console	the operator console

curl -X POST localhost:8000/campaigns \
  -H 'content-type: application/json' \
  -d '{"brief":"eco water bottle for hikers","n":6}'

✅ Tests

37 pytest tests (OFFLINE=1 .venv/bin/python -m pytest) covering: the campaign runs offline end-to-end, the ranking is deterministic + explainable, the top-3 are the three highest scorers, manifest provenance verifies, stored variant PNGs are valid, and the FastAPI surface (health, create/fetch, PNG serving).

🚢 Deploy

Dockerized for Railway (Dockerfile + railway.json, healthcheck /healthz, binds 0.0.0.0:$PORT). The image installs dev dependencies on purpose — Genblaze's OFFLINE mock engine imports pytest at module load, so it is a runtime dependency of the demo path.

🙏 Acknowledgments

Built for the Backblaze Generative Media Hackathon on Genblaze + Backblaze B2. Thanks to the sponsors — Backblaze for B2 object storage and provenance, and GMI Cloud for FLUX image inference (the real-generation provider behind the pipeline):

      

📄 License

MIT © 2026 Edy Cu

---

_{Built by Edy Cu Tjong · GitHub · X · Email}