▶The App ·How it renders 01Roadmap 02Solution 03Value & ROI 04The Room ·The Code Engineering UI ↗

Brand Adaptation Engine

What it does: Swap the hero, rewrite a market's copy, open any stage of the pipeline and change how it composes, then run it. Every millisecond on the chips below is measured on the render happening in front of you — the service reports its own stage timings, and the page replays them. The dwell is capped so the animation stays watchable; the numbers are not.

The pipeline — click any stage to change how it composes
settings apply to every lane
HOW

Why the type is drawn, not generated

The most common way an automated creative pipeline embarrasses a brand is text. It is also the part people most often hand to a model, and that is the wrong instinct — so this is worth being precise about, because a Brand Standards director will ask.

What this engine does

The market's line is set from the brand font's own outlines. The engine reads the TrueType file directly — the character map, the advance widths, the quadratic curves in the glyf table — and converts the copy into vector paths before it ever becomes an image.

Three things follow, and each one is a problem that would otherwise show up after launch:

  • The copy is exactly the copy. It is the string legal approved, drawn glyph by glyph. It cannot drift, misspell, or invent a word.
  • It renders identically everywhere. No font file is shipped and no @font-face is needed, so no viewer's machine substitutes a different face and moves the line breaks.
  • It is measurable. Cap height, line width and block area come from the same outlines that were drawn — which is what lets the legibility rule be a fact rather than an estimate.
Why not let a model render the text
  • Generated text is not typography. An image model produces something shaped like your wordmark. Kerning, tracking and the type scale in the brand book are not reproducible, and near-miss letterforms read as counterfeit.
  • It cannot be trusted to spell. For a legal line or a product name that is disqualifying, and no amount of prompt engineering makes it a guarantee.
  • Non-Latin scripts fail worst. Japanese is where this becomes visible immediately — and it is exactly the market that will notice.
  • It is not editable. Regenerating to fix a comma is absurd. Here, changing the headline is a text edit.
  • It cannot be checked. You cannot measure the cap height of a hallucinated glyph, so the compliance rule that protects the brand has nothing to read.
The division of labour: the model makes the photograph, which is genuinely hard and genuinely creative. Deterministic code sets the type, applies the spec and checks the result — because those need to be exactly right, every time, and exactly right is not what generative models are for.
The outlines, live — the same parser the server composes with
Localization is a font problem before it is a language problem
What this is built on
ASK 01

The Transformation Roadmap

Fifty brands and twenty markets is a thousand brand-market pairs. You cannot sequence a thousand of anything by opinion, and you should not try — the argument in the room is never "is this brand important", it is "whose turn is it". So the sequence comes out of a model the stakeholders can see and argue with. Move the weights and watch Wave 1 change; that is the point. The model is the artefact, not the list it happens to produce today.

What we optimize for
Market weights
Wave 1 gates — pass/fail, not weighted

A high score does not earn a place in Wave 1. These four are binary, because each one is a way for the wave to fail publicly.

Wave size

Wave 1 is deliberately small. Every pair added is another brand kit to normalize, another reviewer to train and another way to be wrong in front of the CMO — and the goal of Wave 1 is not volume, it is a defensible answer to "does this transfer".

Value at stake vs. readiness — all 50 brands
● pilot   ● Wave 1   ● later waves
WAVE 1 ZONE
VALUE AT STAKE →
READINESS →

Readiness is not enthusiasm. It is whether the brand kit is machine-readable, whether the DAM exposes the assets, and whether there is a named owner who can approve. A brand with huge value and no readiness is a Wave 2 problem with a Wave 0 dependency — and saying so early is cheaper than discovering it in month three.

Brands, ranked
Markets, ranked
The six-month journey — click a month
ASK 02

The Co-Innovation Solution

The pilot proved Firefly makes excellent heroes. It did not touch the part that actually costs money. One approved hero has to become an e-commerce tile, a story, a billboard and a shelf card — in six languages, against each market's spec sheet, without breaking brand standards. Today that is a person in a layout file, 500+ times a season. The gap is not generation. It is adaptation. So we build the adaptation engine, and we build it as a bridge the customer's own IT can own.

Where it sits — and what we deliberately did not rebuild

The legacy DAM stays the system of record. We do not migrate it, mirror it or fork it — it gets a read adapter and a write-back of renditions plus compliance metadata. Firefly Services keeps doing hero generation. The engine is the missing middle, and it is roughly two thousand lines with no runtime dependencies, which is the honest answer to "who maintains this when the pod leaves".

What fills the product gap
  • Spec registry. Channel specs are data, not code. A new retail partner is a row, not a release.
  • Subject-aware cropping. A 9:16 cut from a square hero beheads about half of them if you crop from the centre. The engine finds the subject and biases upward, the way a person would.
  • Script-correct typesetting. Market copy is set from real font outlines, so the creative renders identically everywhere with no font shipped and no substitution.
  • A compliance gate that measures pixels. Nine rules, scored on the finished artwork rather than on the brief — and it never auto-approves anything.
How it connects to the legacy DAM
  • Read: approved heroes and brand kits are pulled by reference. The DAM stays the system of record; nothing is migrated.
  • Write-back: each rendition returns with its verdict, its findings, the exact spec it satisfies, and the hero it descends from — so the DAM gains lineage it does not have today.
  • Adapter, not integration. The DAM sits behind the same interface as everything else. Swapping the object store on this deployment took one file and no changes anywhere else — that is the maintenance claim, tested.
  • No new source of truth. If the engine is switched off tomorrow, the DAM is unchanged and every asset it produced is still there.
ASK 03

Value & ROI

Productivity is the easiest number to produce and the easiest for a board to discount, because "we saved hours" invites the question "so where are they". The metrics below are the ones that survive that question: they measure whether the company can now reach markets it could not afford to serve properly, whether it can commit to dates, and whether the risk of a brand or legal defect reaching the public has gone down. Cost avoidance is the floor, not the case.

Model the case — your assumptions

Every input is an assumption the CMO's finance partner will challenge, so they are all exposed rather than baked in. The defensible version of this slide is the one where they change a number and the conclusion holds.

Outcome
Where the hours actually go

The claim is not headcount reduction. It is that adaptation stops consuming the people you hired to originate — which is the version the Creative directors will accept and the board can still bank.

The board scorecard — six metrics beyond productivity
And the honest half — what would tell me this is failing

A transformation dashboard that can only go up is a marketing asset, not an instrument. These are the leading indicators I would put in front of the same board, and I would rather name them now than be asked for them in month five.

04

The Room

Five people with five different reasons to say no. Each concern below is legitimate, and each one is answered by something that is already running rather than by a promise on a slide — click a stakeholder to see the answer and the proof.

SRC

The code

All of it, served by the thing it builds. No runtime dependencies — the YAML reader, the PNG codec, the TrueType parser and the compositor are in this repository, on top of primitives the Workers runtime already has. That is the honest answer to "who maintains this after the pod leaves".