Video Reasoning Engine

The architecture

Today's marketing workflow is broken into disconnected pieces, market research on audience and channel, creative decisions, production, and publishing and analysis, each typically living in a separate tool. Alex Mashrabov's stated long-term goal is "to build a video reasoning engine which closes everything in the loop."¹

The build is staged in three parts. It starts with generation, the current core capability. It then adds visual-understanding modules, a semantic layer giving the system awareness of its own outputs, which enables visual consistency throughout a single video and end-to-end video creation inside one platform rather than across several. The final stage is publishing, analyzing, and proactively suggesting next steps, which requires real awareness of a brand's history and a given campaign's past performance, effectively a persistent model of the brand and its results rather than a stateless generation tool.

Why it gates outcome-based pricing

The reasoning engine is explicitly framed as the prerequisite for pricing on outcomes rather than on usage: "this is not possible today because the models don't have the reasoning engine to really connect the dots between the data, the visuals, the generations." Only once the loop closes end to end can a platform meaningfully claim to drive an outcome, and only then can it charge for the outcome itself rather than for each individual generation.

Why it matters

The architecture reframes a generation tool into something closer to a marketing operating system: the durable value is not producing one polished clip but owning the full research-to-production-to-measurement loop and the brand context that accumulates inside it over time, a defensibility argument resting on the idea that accumulated context is sticky in a way raw model access is not. It is also the concrete mechanism behind outcome-based pricing broadly, since a system can only sell outcomes once it can perceive, attribute, and iterate toward them. And the visual-understanding layer, giving a generator awareness of its own prior output, is a narrow, applied version of a much larger idea in AI research: models that carry some persistent, structured awareness of the world they are generating into, rather than treating every output as a stateless one-off.

Open questions

This is a stated roadmap rather than a fully shipped capability, with outcome-based pricing described as a goal for later in the decade and the hardest components, reliable attribution and genuinely proactive suggestion, dependent on data that social platforms do not yet govern clearly for AI-generated content. Closing the loop also requires clients to share their own performance data back into the system, which is being piloted only with clients comfortable doing so, a real adoption gate rather than a given.