The context
PromoGenAI turns a product description into a finished video ad. Not a storyboard, not a draft for an editor to rescue: strategy, script, scene composition and final render, all handled by an agentic engine. It exists because performance creative is the bottleneck for small teams. The ad you can afford is usually the one you never make.
The idea had a precursor. At Markhorverse I built a prompt-to-video generator: enter a prompt, get scenes, scripts and a compiled short-form video. It proved the mechanism and exposed the real problem. Raw generation is not the product; judgment is. PromoGenAI is that lesson built properly. Before it renders anything, the system researches the brand and proposes the persuasive angle.
The constraint
Two constraints shaped the build, and neither of them was creative. The first belongs to the customer: an ad has to cost less than the campaign it serves, so the price of a single run is a product decision long before it is an infrastructure one.
The second belongs to the product. It is independent, with no agency production line standing behind it, so every component has to be boring enough to run unattended and every piece of complexity has to buy its place. Independent products rarely die of missing features. They die of maintenance debt.
The architecture
React and TypeScript on Vercel; FastAPI and Python on GCP Cloud Run; PostgreSQL underneath. The interesting fifth of the codebase is the pipeline: a multi-stage agentic chain that orchestrates Claude for strategy and script, GPT-image-2 for scene composition, and Seedance for the final video render. Each stage produces a validated artifact before the next one runs. Strategy, then script, then scenes, then render.
Around it sits the unglamorous majority: Stripe credit-based billing, Google OAuth, deployment through GitHub Actions. That part is what makes it a product rather than a demo.
The hard part
Chained generative stages multiply failure modes. A weak strategy makes a weak script makes an expensive, unusable render, and video generation is the costly stage, so mistakes have to be caught before the money is spent. The pipeline validates each stage’s output against the brief before the next stage spends anything.
That is also why billing is credit-based. Cost per run is a product feature, not an infrastructure detail: it has to be as legible to the user as it is to me, and the architecture is shaped around keeping it inside a hard ceiling.
Where it is now
Live at promogenai.com. The engine researches a brand, proposes ad angles and produces finished video creative; billing, auth and deployment run themselves.
The problem it was built against has not moved. Performance creative is still the bottleneck in advertising, and the teams that feel it hardest are the ones with no production line to put behind it. An independent product can serve that gap properly, provided what gets engineered is the loop rather than the feature list.
- React
- TypeScript
- FastAPI
- Python
- PostgreSQL
- Claude
- GPT-image-2
- Seedance
- Stripe
- GCP Cloud Run
- Vercel