Most agency case studies are about somebody else’s project, written after the fact, with the difficult parts removed. This one is mine. VeoStudio is an AI Makers product: I designed it, I wrote it, I run it in production, and I pay the bills when a design decision turns out to be wrong. It is the clearest answer I can give to “what can you actually build?”
So here is the build, honestly, including the part I got wrong and had to redo this month.
What VeoStudio does
You start with a script, or a brief that becomes one. VeoStudio breaks it into scenes, generates each scene as a video shot, and assembles the shots into a single continuous film with characters that stay recognisable from one shot to the next. The generation runs on Google Veo 3.1; the scene writing runs on Gemini 3.7 Flash. Around that sits everything a person actually needs to finish a video: a director chat, templates, a character library, a timeline, and stitching.
The interesting engineering is not the model call. Anyone can call Veo. The interesting part is everything between a paragraph of text and a finished film:
- Turning prose into shots. A script is not a list of prompts. Scene generation has to decide where the cuts go, what each shot needs to contain, and what carries over from the previous one — and it has to be fast and cheap enough to re-run when the writer changes their mind, which is why it runs on Gemini 3.7 Flash rather than a heavier model.
- Character continuity. Generative video has no memory between calls. If the same person is meant to appear in nine shots, that has to be enforced from outside the model, with reference images carried through the pipeline.
- Long jobs that fail halfway. Video generation is a long-running operation, not a request. Every shot is a submit, then a poll, then a download that has to be re-hosted before anyone can watch it — and each of those steps can fail on its own, hours after the user closed the tab.
- Constraints that are not yours. Shots are 4, 6 or 8 seconds — nothing in between. 1080p needs the full 8 seconds. Frame-to-frame continuity needs exactly 8. Reference images are allowed on some model tiers and refused outright on others. None of that is negotiable, so the interface has to make it obvious rather than letting someone discover it in an error message.
That is the honest shape of an AI product in 2026. The model is a component. The product is the orchestration around it, and that is ordinary, careful software engineering.
The update: bring your own key
VeoStudio originally ran on one billing model — tokens. You bought a balance, generation spent it, I paid Google behind the scenes. It is the standard approach and it has a standard problem: any resale meter is either too expensive for heavy users or too cheap to sustain, and there is no setting that is fair to both.
So this month I shipped the alternative. Pro members can now paste their own Google AI Studio API key into the studio, and from that point every Google job — video, images, frames, music — runs on their own Google account at Google’s own list prices and spends zero tokens here. Their token balance simply sits untouched, and the monthly Pro allowance still lands.
The parts that turned out to matter more than the feature itself:
- Validate before storing. The key is checked with Google before it is saved, so a key that will not work never becomes a mystery failure three days later. After it saves, only the last four characters are ever displayed.
- Never let the key reach a browser. Google’s finished video is served from a URL that only works with the key attached — fetch it without one and you get a 403. That means the download has to happen server-side and be re-uploaded to our own storage. Handing that URL to a browser would have been the fastest implementation and a straightforward way to leak a customer’s credentials.
- Check the shot before submitting it. The Gemini API and Vertex AI speak similar but not identical dialects, and a few options are silently unavailable on a personal key. VeoStudio checks each shot against what that key can actually do and names the offending setting up front, rather than letting Google reject the job after the user has waited.
- Make failure boring. If Google refuses a key or you hit a quota on your own project, generation falls back to tokens instead of stopping. “Use tokens instead” is one click, and the key stays saved but dormant.
Two things stay on tokens either way, and I would rather say so plainly than bury it: the non-Google video engines — Runway, Kling, MiniMax, Sora — cannot be paid for with a Google key, and character voices run on the studio account. Everything that was already included with Pro stays included.
The commercial lesson generalises well beyond video. Charging for the thing you actually built, and letting customers pay the infrastructure provider directly, removes an entire category of argument from your support inbox. I now raise it in most discovery calls where someone is planning to resell model usage.
Need this built and shipped for a real business?
If you read this guide for a client or for your own company instead of yourself, we build the production version. Custom AI software development from €5K WhatsApp bots to €150K platforms — fixed price, weekly demos, your code is yours.
Or build your own AI system piece by piece and send the design in for a written quote →
What this has to do with hiring me
A portfolio of client logos tells you who was willing to sign. A product you can go and use tells you rather more. VeoStudio is production software with real users, real bills, an actual support inbox and a public help centre, and the same practices go into client work:
- Long-running jobs are designed to be resumable, because in the real world people close the tab.
- Credentials are handled as if they will leak, because the cheap implementation almost always would have.
- Constraints are surfaced in the interface, not discovered in an error.
- Pricing is a design decision, not something bolted on at the end.
If you are commissioning something in this shape — an AI feature inside a real product, with jobs that take minutes, credentials that belong to your customers, and a bill that has to make sense — that is exactly the work described on the custom AI development page, and the cost estimator will give you a range for it in about thirty seconds. Other things I have shipped are on the work page.
If you just want to make a video
Go and use it. VeoStudio has the pricing, the help centre and the bring-your-own-key guide, and none of it needs to go through me.
Common questions
What is VeoStudio?
A web studio for making multi-scene AI video. You start from a script or a brief, it is broken into scenes, each scene is generated as a shot on Google Veo 3.1, and the shots are assembled into one film. It is a product I build and run at AI Makers.
Which AI models does VeoStudio use?
Video is generated on Google Veo 3.1 — the draft, fast and quality tiers map to Veo 3.1 Lite, Veo 3.1 Fast and Veo 3.1 respectively. Script and scene writing runs on Gemini 3.7 Flash. Non-Google engines including Runway, Kling, MiniMax and Sora are available for individual scenes.
What does bring-your-own-key mean in VeoStudio?
Pro members can paste their own Google AI Studio API key into the studio. Every Google job then runs on their own Google account at Google list prices and spends zero VeoStudio tokens. The key is validated with Google before it is stored, only the last four characters are ever shown again, and it can be switched off at any time to go back to tokens.
Why would a product let customers pay Google directly instead of paying you?
Because the alternative was worse. Metered resale means heavy users feel overcharged and light users subsidise them, and every support conversation becomes an argument about the meter. Letting heavy users bring their own key removes the argument entirely, and the subscription then buys the thing that is actually mine — the studio, the scene engine, the timeline, the character library.