Upload a brand book, name an occasion, and a crew of specialist AI agents research the market, set strategy, write the copy, generate real visuals and a storyboard, localize it, and then look at their own images and critique them before a human signs off. Not a mockup — a working multi-agent system with guardrails, tracing, cost control and a human-in-the-loop.
Marketing teams spend days turning a brief into research, copy, and creative. Cortex compresses that into one orchestrated run and hands back an editable, brand-checked kit: strategy, copy, a hero visual, a poster, a social carousel, a storyboard with a cut video, and a localized version — each asset scored for brand fit by a vision model that can send it back to be regenerated.
Drop in a brand book (PDF or image). A multimodal agent extracts the palette, voice, and do/don'ts, so everything downstream stays on-brand.
"New Year campaign for our EV scooter." Agents research the market, the moment, and the audience before a single word is written.
Copy, hero, poster, carousel, a storyboard and a stitched video — real generated assets, not prompts.
A vision agent looks at every image, scores brand-fit and quality, and regenerates what fails. Then you approve, edit, or reject.
Ten specialist agents on a LangGraph state machine, each resilient (one failure degrades the run, never crashes it). Every model call is routed by cost tier, metered, and traced. A human gate can pause the whole graph and resume it.
Prompt-injection neutralization, PII redaction, output moderation, publish-gating, rate limits. An adversarial red-team test suite proves it.
Short-term checkpointer, long-term brand vectors, episodic run history. Research is grounded in the uploaded brand, with citations.
Per-agent tracing (latency, tokens, USD), model-tier routing, prompt caching. You see exactly what each step cost.
Golden-set evals, an LLM judge, and 52 automated tests including concurrency, fuzz, and the visual-QA regenerate loop.
A real run for the Vespa Elettrica (New Year, Italy). Press play to replay the agents and see the finished kit — copy, generated visuals, storyboard, Italian localization, and the self-QA scores.
Upload a brand book and describe a campaign. The real backend runs the full multi-agent pipeline on live models. This lite preview uses free-tier models on a sleeping free server, so expect a ~40s cold start, a few minutes per run, and a small daily cap. Wired to a paid model from the open-source code, it is far faster and sharper.