Salvo

• 

  `OFFLINE=1 pytest -v` real output: 19 passed in 0.48s.

• 

  One real API call: POST /campaigns returns the whole batch — manifest_verified: true, ranked top-3, and winner reasons.

• 

  Close frame: 19 tests green · deterministic explainable ranking · verified provenance — with the OFFLINE/mock-images honesty note.

• 

  "One ad is a guess — a salvo is a decision": brief in → N scored variants out → best three ship (brand hook frame).

• 

  Live `/console` full view: brief + N in, verified manifest hash out, top-3 and full batch below (OFFLINE demo, mock placeholder images).

• 

  Live /console top-3 grid — each variant's score + full reason breakdown; engagement is a content-hash stand-in, not real click data.

💡 Inspiration

Generative tools make it trivial to produce one ad. The real work is producing many and knowing which to ship — and being able to explain why. Most "generate N variants" tools hand you a wall of images with no ranking, no provenance, and no reproducibility. We wanted a small, honest pipeline that turns a brief into a batch, scores every variant on transparent signals, and surfaces the best three — with the reasoning shown, not hidden.

⚙️ What it does

Give Salvo a creative brief and a count N. It fans out N ad-image variants through one real Genblaze Pipeline, stores each variant with a verified provenance manifest, scores them with an explainable, deterministic ranking, and surfaces the top 3 to ship — each score accompanied by a plain-English breakdown of exactly how it was reached.

Every variant is scored on three transparent signals:

Honesty note. The engagement index is not real click/CTR data — it's a deterministic stand-in seeded from the content hash so the ranking is reproducible offline, and every reason string says so.

🏗️ How we built it

• Engine — Genblaze: a real Pipeline.astream(max_concurrency=N) batch fan-out, one step per variant, stored through an ObjectStorageSink over a LocalDirBackend (a real subclass of Genblaze's documented StorageBackend interface), with provenance read back via read_manifest(verify=True).
• Storage — Backblaze B2: OFFLINE uses the local backend; setting B2 credentials switches storage to a real Backblaze B2 bucket via Genblaze's S3StorageBackend (auto-detected at startup).
• App: a FastAPI backend with an operator console (/console), a dependency-free raw-PNG encoder (no Pillow/ffmpeg), and an in-memory campaign store. Packaged with uv, containerized with Docker, deployed on Railway.
• Ranking: deterministic and explainable — same brief + variant always yields the same score and the same reasons.

🧗 Challenges we ran into

• Explainability without a real model in the loop. We wanted honest, reproducible scores in an offline demo, so we designed the ranking around transparent, deterministic signals and made the "engagement" signal explicitly a content-hash stand-in — labeled as such everywhere — rather than fabricating fake CTR numbers.
• A real batch pipeline, not a loop of fakes. The variants flow through a genuine Genblaze Pipeline fan-out and are provenance-verified, so the "batch factory" claim is backed by real SDK code paths, not a for-loop.
• Zero-dependency images. To keep the OFFLINE path fast and portable (no Pillow, no ffmpeg), we wrote a tiny raw-PNG encoder for the mock variant images.

🏆 Accomplishments that we're proud of

• A working, honest batch-and-rank pipeline on real Genblaze code paths, deployed live, in a focused build.
• Explainable ranking: every score ships with its reasoning, and the one non-real signal is labeled as such — no fabricated metrics presented as real.
• 37 tests, OFFLINE-by-default (zero credentials), Dockerized and live on Railway.

📚 What we learned

• The valuable part of "generate N" isn't the generation — it's the selection, and selection is only trustworthy if it's explainable and reproducible.
• Genblaze's Pipeline + StorageBackend abstractions make a real batch pipeline with verified provenance small to build.

🚀 What's next for Salvo

• Swap the mock image provider for real generation (GMI FLUX / DALL·E) behind the same pipeline — the seam already exists.
• Real Backblaze B2 storage + a promotion-on-storage-event delivery step (approved variant → webhook).
• Richer ranking signals (brand-safety, visual diversity) and a live A/B feedback loop to replace the seeded engagement stand-in.

Built With

• amazon-s3-api
• backblaze-b2
• docker
• fastapi
• genblaze
• python
• railway
• uvicorn