Metropolis.
From a single AI-generated still, eleven agents orchestrate strategy, creative, copy, motion and compliance — turning weeks of variant production into minutes.
An agentic platform for turning a single still into the full story.
Metropolis lets creative teams configure a campaign — pick the audience persona, the brand, the channels — then generate a full set of broadcast-ready ad variants from a single hero still. Output is sized and timed for Instagram, Meta, YouTube, and display, without anyone manually reformatting a thing.
A WYSIWYG canvas gives creative directors a single place to preview every variant, swap imagery by typing what they want, trim motion, regenerate clips, and approve the final cut. Everything stays inside a brand-safe environment — colour palettes, logo zones, and exclusion rules are enforced before anything reaches review.
Eleven specialised agents handle the work behind the scenes — strategy, creative direction, motion, copy, QA, brand compliance — running in parallel where they can. What used to take weeks of manual variant production now finishes in minutes, with creative directors stepping in only at the decision points that need a human eye.
A canvas creative directors actually ship from.
Every variant in one place. Swap imagery by typing what you want — “cleaner sky”, “dusk light”, “hero in shot” — and the canvas finds it, drops it in, and re-runs the frame. Scrub the timeline, trim or regenerate any motion clip, and reformat for every channel without leaving the surface.
At each creative checkpoint the pipeline pauses for the director — Approve, Redo with feedback, or Override. Decisions sit with the people responsible for the work, not with a background job.
Cut prompts by 80% before Gemini ever sees them.
A pgvector dual-embedding retrieval layer narrows candidate pools by ~95% before the LLM picks. Blended ranking — 60% text, 40% visual — tuned on real campaign content with a graceful text-only fallback when CLIP embeddings miss. The cost saving compounds across the pipeline; what was $0.40 per variant in tokens is now closer to $0.08.
- image library~12,400 images
- semantic shortlist~640 ranked matches
- final selection by Gemini40+ images
Custom agent orchestrator
Designed and built the orchestration layer that runs the eleven agents end-to-end. New agents — Insights, QA, Brand Compliance — slot in as the platform grows without rewriting anything upstream, so the system keeps pace with how the brief evolves.
Smarter image retrieval
Built the search layer that finds the right hero image for a given persona and brand. Narrows candidate pools to a sharp shortlist before the LLM picks — keeping selections relevant and prompts cheap, even as the image library scales into the tens of thousands.
Multi-provider video generation
Wired VEO 3.1 and Runway Gen 4.5 behind a single interface so the team can swap providers per project, balance cost against quality, and adopt new models as they ship — without touching the rest of the pipeline.
WYSIWYG Studio
Lead engineer on the editing surface creative directors actually ship from. Type to swap any image; scrub the timeline and trim or regenerate any clip; reformat for every channel without leaving the canvas.
Human-in-the-loop review gates
Built the review checkpoints that pause the pipeline at every creative moment that needs a human eye. Directors get clean Approve, Redo-with-feedback, and Override controls — decisions stay with the people responsible for the work.
