AdTech Creator Monetization Platform

Why a manual pipeline could not scale

Every campaign moved through four hands-on stages: someone watched the submitted video to check it met the brief, someone rendered the branded overlay onto it, someone uploaded the finished cut to each platform, and someone arranged the creator's payment. None of those steps is difficult on its own. The problem is that they run in sequence and each one waits on a person, so a delay anywhere stalls everything behind it — which is how a pipeline whose actual work totals a few minutes ends up measured in weeks.

That cost compounds in a creator marketplace specifically, because both sides feel it. Brands wait to see their campaign go live; creators wait to be paid. A platform that is slow at either loses the participants it depends on, and at ten creators that attrition is not a statistic — it is a named person deciding not to come back.

The constraints that shaped the build

Three things narrowed the design. The first was funding. A $150K-funded startup does not have the runway to staff a review team while it is still finding product-market fit, so automation was not an optimisation applied later — it was the only version of the product that could exist at that stage. Whatever got built had to run without a human in the loop for the ordinary case.

The second was that video is heavy. Inspecting a file, re-encoding it with an overlay, and pushing it to two external platforms is not request-scoped work, and none of it fits inside the lifetime of an HTTP request. The pipeline had to be asynchronous by design rather than made asynchronous once it started timing out.

The third was that money moves at the end of it. Paying a creator is not a notification, it is a transfer to a third party — which brings onboarding, identity verification, and payout mechanics into scope alongside the media work, and makes the last stage of the pipeline the one with the least tolerance for being approximately right.

How the pipeline works

Validation runs first, using Gemini to inspect the submitted video against the campaign's requirements. This is the stage that most clearly needed a model rather than a rule: whether a video matches a brief is a judgment about content, not a property you can read from a file header. Putting it at the front also means the expensive stages never run on a video that was never going to qualify.

Rendering is FFmpeg applying the branded overlay to an approved video. It is deterministic work with a predictable failure surface — codecs, aspect ratios, durations a platform will not accept — which makes it the right place to be strict. A rendering problem caught here costs a retry; the same problem discovered at upload costs a rejection and a round trip.

Publishing pushes the finished video directly to TikTok and YouTube. This stage is less about video than about two third-party integrations, each with its own API, its own authorisation model, and its own definition of a valid upload. They fail independently, which means they have to be retried independently — a successful TikTok upload must not be undone by a YouTube error.

Payouts run through Stripe Connect, which is what makes this a marketplace rather than a publishing tool. Moving money to third parties brings creator onboarding and transfer mechanics that would otherwise be a project of their own, and Connect absorbs most of that.

A Next.js application ties the four stages together, giving both sides of the marketplace a single place to submit work and see where a video currently sits in the pipeline.

Why the numbers matter

Weeks to minutes is the headline, and it is a change in kind rather than degree. A pipeline measured in weeks can only serve campaigns planned weeks ahead. One measured in minutes can respond to something happening now — which is the entire premise of short-form content, and the reason the same platform with the same creators becomes capable of work it simply could not take on before.

Ten creators and one hundred videos are deliberately unglamorous figures, and they are the honest ones to state. They are what the pipeline actually carried. At that stage the useful question is not how large the numbers are but whether the system runs unattended, and a hundred videos through an automated pipeline is a hundred campaigns nobody had to shepherd by hand.