ABC Affiliate Network was confidential work. NDA prevents disclosing the product surface, the asset taxonomy, or the affiliate-facing flows. What follows is the role and the architectural shape — without revealing anything covered by the agreement.
What it is
A platform for managing brand assets across the network of TV station affiliates that carry ABC content. Not a glamorous surface — the kind of internal tool that, when it's working, no one notices. When it's broken, an entire affiliate operation can't air the spot they're contractually obligated to.
The shape of the problem
Enterprise asset management has a few hard parts that all meet in the same place: taxonomy (how do you classify thousands of pieces of content so the right affiliate gets the right thing?), permissions (which affiliates can pull which assets, under what licensing window?), and operational surface (how do non-technical operators get their work done without a training course?).
I worked on the third — the operational surface. Front-end engineering plus UX strategy. The taxonomy and licensing logic lived in the backend; the platform's job was to expose them in a way that affiliate ops teams could actually use under deadline pressure.
What I learned
Enterprise tools fail because they're optimized for the wrong audience. The team building the tool reads the schema; the people using the tool don't. The interface either translates the schema into operational language, or it forces the user to learn the schema. The second path is what produces internal tools that everyone hates.
ABC Affiliate Network worked because we treated the operational language as the source of truth and translated to the schema, not the other way. Affiliate ops doesn't care about asset_type_id; they care about "the eight o'clock spot for Tuesday." Get the surface right, and the platform becomes the way work gets done. Get it wrong, and the spreadsheet wins.
Why this matters for AI systems
Production AI ops surfaces are exactly this problem. The model researcher reads the trace schema; the on-call engineer responding to a 3am incident does not. If the eval dashboard is a thin shell over the trace JSON, it fails the same way ABC's first internal tool would have failed — operators don't have time to translate, and the tool stops getting used.
The discipline that worked at Disney is the discipline that works for AI ops: the operational language is the source of truth, the schema translates to it, and the interface is the only thing your operators ever have to look at. Get that right and the system becomes the path of least resistance. Skip it and you've shipped infrastructure no one will operate.