SoMe Studio
From virtual model to running ad: AI creatives that prove themselves on Meta and Google.
SoMe Studio lets entrepreneurs and agencies define a virtual model (persona) once and then produce endless ad variants — image, copy and video. Variants go live on Meta and Google with a small daily budget; a daily rebalancer automatically scales the winners.
9
Inngest workflows
36
Tables, all with RLS
30 min
Metrics sync cycle
80/20
Budget to proven winners
What it does
Persona Designer
Reference photos plus structured input become a reusable model sheet with a canonical description — so the same virtual model returns consistently in every campaign.
Batch generation
One creative brief ('model B with product D on the beach') fans out to multiple personas, image models and photo styles at once; approve, reject or regenerate per variant.
Test-and-optimise loop
Metrics sync every thirty minutes; the daily rebalancer scores on CTR and CPA, gives the top-2 variants 80% of the budget and pauses the rest.
Credits & Stripe
Append-only credit ledger with atomic deduction per AI call and per euro of ad spend; packages via Stripe Checkout with idempotent webhooks.
Video productions
From AI news videos (script → voice → shots → FFmpeg render with lower thirds and end card) to avatar briefing videos and vertical scene ads with a branded packshot.
Model catalogue as data
Image and video models (Flux, GPT Image, Kling, Veo, Seedance …) live in the database as data, each with its own credit price and margin — releasable per workspace, without a deploy.
How we built it
Next.js with Server Actions, Supabase Postgres with row-level security on every workspace table, and Inngest for all heavy lifting: nine workflows for generation, launch, metrics sync and budget rebalancing.
The ads side sits behind a provider-agnostic adapter — one interface, everything in euro cents — so switching ad platforms never rewrites a workflow.
Three-tier LLM routing (cheap/standard/premium via OpenRouter) keeps model choice in one place, and video renders with ffmpeg on Vercel — bundled fonts included, because serverless functions ship none.
Why we built it
Ad creative does not scale: every new angle needs a shoot with a model, product and location — while ad platforms want volume in variants to test against each other.
SoMe Studio flips it: define a model once, then produce variants at marginal cost — with the test loop built straight in. The AI news stills on this page come directly out of the pipeline.
Need something like this in your company?