Building and running a live multi-model AI platform

AI capability is fragmented. Different models are better at different things, and each sits behind its own API with its own request format, failure modes, latency profile and pricing. Using several of them well is a genuine engineering problem.

PotatoAIHub was built to solve that in one place — chat, image generation and video generation across multiple models behind a single interface.

It is a real product with real users, which means it also has to deal with everything a demo does not: cost per request, requests that take minutes rather than milliseconds, providers that fail, and the privacy expectations that come with people sending you their prompts.

The central design question was how to add a new model without the application needing to know anything new. Provider APIs differ enough that naive integration produces a codebase where every model is a special case.

Generation workloads also break the assumptions of a normal web request. Image and video generation take far longer than a request should be held open for, so the work had to be modelled asynchronously from the start rather than retrofitted.

Model providers sit behind a common internal interface, so the application works against one abstraction and each provider's differences stay contained. Adding a model is an integration, not a refactor.

Generation runs through queued background processing with results delivered when ready, which keeps the interface responsive regardless of how long a given model takes.

The platform runs on AWS and is operated as a live product — deployment, monitoring, cost control, and the ongoing work of keeping something running rather than shipping it once.

PotatoAIHub is live at potatoaihub.com, serving multi-model chat, image and video generation in production.

It is the clearest available evidence for the AI work described on this site: the same integration patterns, async processing and cost control a client's AI feature will need, already built and running.

It also directly informs the advice given to clients about AI — including where AI genuinely helps, and where it is an expensive way to solve a problem that had a simpler answer.