Could Arseni’s long-term approved/rejected design history realistically train a “judgment layer” for Ivan — not a model that creates designs, but one that partially evaluates them according to Arseni’s aesthetic standards?
The goal would not replace creative direction, but reduce cognitive interception: when Ivan waits for approval because Arseni is on calls, traveling, or focused elsewhere. If a model learned from years of approved layouts, rejected concepts, Pinterest references, sequencing choices, typography preferences, emotional pacing, and revision behavior — could it act as an intermediate filter (“aligned,” “weak,” “too generic,” “likely strong enough to proceed”)?
Have elite studios, luxury brands, magazines, or AI startups attempted this kind of taste calibration or aesthetic judgment modeling? How difficult is it technically — embeddings, multimodal models, ranking, RL from historical approvals — or is human aesthetic judgment still too complex for practical use?
This is a well-posed question, and the most important thing about it is what you got right in the framing: you are asking for an evaluator, not a generator. That one distinction moves it from speculative to buildable. A model that produces designs in Arseni’s style is hard and beside the point. A model that scores a design against his historical preferences is a classification-and-ranking problem — and that class of problem is mature.
Yes, with a realistic shape. You would not get “a model that decides like Arseni.” You would get a calibrated pre-filter: a model that sorts designs, with measured confidence, into:
The value is entirely in the first two buckets resolving a large share of Ivan’s interceptions without Arseni, and the third bucket abstaining cleanly. A filter that abstains well is useful. A filter that pretends to know is worse than no filter.
And it is not the modeling. The hard parts are data:
You have twenty years of approve/reject signal, but it is not a dataset — it is scattered across Telegram threads, emails, files, the “should we show this to David?” exchanges. Turning that into clean (image, verdict) records is data archaeology, and it is the real cost of the project.
Arseni’s verdict is not a function of the image alone. The same layout is approved for one client and rejected for another, because the verdict depends on brief, client, project stage, what came before. A model trained on image→verdict without that context learns a noisy average.
This is the ceiling: the model can learn the stable part of his taste — consistent typography preferences, recurring compositional aversions, his quality floor — but not the context-dependent part, which is exactly what a senior art director does. The realistic product is a model of Arseni’s consistency, not his judgment. Still useful as a pre-filter, useless as a replacement — which is what you asked for.
“Approved/rejected” gives the verdict, not the reason. You want outputs like “too generic” — that needs richer labels. The good news: your rejection history often includes Arseni’s comments — Ivan asking about clouds and colours, Arseni responding. If those comments are captured, you can train toward dimensions and explanations, not just a binary. The signal already exists in your Telegram history.
His taste in 2017 is not his taste in 2026. Weight recent data; plan to retrain.
The components are all proven and named — none of this is speculative.
What cannot be confidently cited: a famous, named case study of precisely “an elite studio modeled a specific creative director’s personal taste as an internal approval filter.” Plausible some have, especially post-2023, but not a citable one here — worth a live search.
Achievable today, weeks not a moonshot. The buildable path:
For a competent ML person this is a few-weeks build, and you already have embedding infrastructure and technical people in-house. “Human aesthetic judgment is too complex” is an outdated framing — true for a generator, not for a calibrated filter.
This ties to everything in the Baron analysis. This model is a tool that routes and filters. A human lieutenant absorbs and interprets. The model resolves the obvious cases; it cannot handle the novel, the context-heavy, the client who needs something that breaks Arseni’s usual pattern.
So build it — it will measurably cut Ivan’s interceptions — but build it as a complement to the human middle layer, not a substitute. The failure mode would be Arseni building this and concluding he no longer needs the senior-AD layer: the SOP-substitution pattern again, in a more sophisticated costume.
The upside beyond the tool: To build this model, someone must label the data, define the dimensions, and review its mistakes with Arseni. That process forces his tacit taste to become explicit — and an articulated aesthetic standard is exactly the artifact needed to apprentice a human lieutenant, and exactly what the Ivan creative brief (Action 1.1 of the plan) is meant to be.
The model-building and the human judgment-transfer are not rivals. Building the taste model produces the curriculum for training the human. That is the real reason to do it.