How We Feed Raw Brand Assets Into an AI and Get Back a Ranked Shortlist in 30 Minutes14
The 30-Minute Pipeline: How We Turn Raw Brand Assets into a Ranked Shortlist
You have a brand. You have 400 product images, 120 color swatches, 30 typography samples, 50 brand voice notes, a logo file, a style guide PDF, and a folder of "inspiration shots" that nobody has opened in six months. You want to know, in 30 minutes, which of your 2,400 SKU-level product concepts are the strongest candidates for the next campaign.
You don't want a report. You don't want a deck. You don't want a meeting. You want a ranked shortlist.
That is the problem this pipeline solves. And it is not a trick — it is a small, well-ordered stack of AI components doing exactly the job they are good at.
What "raw brand assets" actually means
Before we talk about the pipeline, let's be precise about the input. Raw brand assets are the unstructured, semi-structured, and sometimes just plain messy artifacts that define a brand. In our stack, they come in four families:
Visual assets. Product photos, hero shots, swatches, mood boards, logo variants, UI screenshots.
Textual assets. Brand voice guides, tone documents, taglines, product descriptions, review corpora, press kits.
Structural assets. SKU catalogs, P&L data, inventory depth, pricing tiers, category trees, channel mix.
Relational assets. Which products were co-purchased, which were bundled, which drove repeat purchase, which were returned.
The last family is the one most pipelines forget. Brand is not just what you say about a product. It is how the market behaves around it. A product that is on-brand but rarely bought is a weaker candidate than one that is slightly off-brand but has a loyal audience. The pipeline treats both signals as first-class citizens.
The raw input set we work with is typically between 800 and 5,000 files. Total size between 2 and 40 GB. Heterogeneous: JPEG, PNG, PDF, CSV, JSON, XLSX, and the occasional scanned image that needs OCR before it is useful.
The goal: in 30 minutes, end with a ranked list of 20 to 50 SKUs, each with a score, a one-line rationale, and a set of evidence pointers back to the original assets.
The four stages of the pipeline
The pipeline has four stages. Each one is a small, testable unit. Each one has a clear input contract and output contract. None of them are "magic."
[Raw Assets]
│
▼
[Stage 1] Ingest & Normalize ──► A clean, typed asset graph
│
▼
[Stage 2] Embed & Index ──► A hybrid search index
│
▼
[Stage 3] Score & Rank ─► A ranked, explainable shortlist
│
▼
[Stage 4] Package & Deliver ─► A human-readable artifactEach stage runs in parallel where possible. Total wall-clock time, end to end, is 22 to 28 minutes on a mid-size asset set.
Stage 1: Ingest and normalize
The first job is to make everything addressable. A JPEG and a CSV should live in the same graph. A brand voice note and a P&L row should be queryable side by side.
We build an asset graph where nodes are individual assets and edges are relationships. A product photo is linked to its SKU. A SKU is linked to its category, its price point, its channel mix, its review corpus. A brand voice note is linked to the product families it was written for.
Normalization does three things:
Type inference. A file that looks like an image but is actually a scanned table gets OCR'd and reclassified.
Deduplication. Near-duplicate hero shots are collapsed. A product photographed three times in the same studio setup becomes one node with three provenance pointers.
Metadata completion. Missing fields get inferred. A product photo without a SKU gets matched to a SKU by visual similarity against the catalog. A brand voice note without a category gets one by matching its language to the category tree.
The output is a typed graph. Every node has a type, a set of tags, a provenance pointer, and a quality score. Quality score matters. A hero shot at 4000 pixels is not the same evidence as one at 800 pixels. The pipeline weights evidence by quality.
This stage is deterministic. No model is involved. It is a parser, a deduplicator, and a graph builder. It takes about four minutes for a 5,000-file set.
Stage 2: Embed and index
Now the graph gets turned into something searchable. This is where the AI does its first real job.
We use three embedding models in parallel, each optimized for a different modality:
A vision-language model for images. It encodes the visual content of each asset into a 1024-dimensional vector.
A text embedding model for textual assets. Brand voice notes, taglines, descriptions, review corpora.
A tabular encoder for structured data. Pricing, P&L, inventory, channel mix. This is a small transformer over the structured fields.
Each asset gets three vectors. The asset graph stays intact — the vectors are indexed, but the graph is the source of truth.
We build a hybrid index:
A vector index for similarity search.
A filtered index for type, category, channel, price tier, and quality score.
A relational index so you can query "all products co-purchased with SKU X" as a first-class filter.
The index is built in about six minutes. It is searchable in sub-second latency.
The trick in this stage is that we don't collapse modalities into one vector space. We keep them separate and let the scoring stage do the cross-modal reasoning. A good product concept is not "similar to the brand voice note." It is "visually consistent with the hero shots, linguistically aligned with the brand voice, and commercially consistent with the price tier and channel mix." That is a four-way alignment, and it is easier to reason about when the embeddings are kept distinct.
Stage 3: Score and rank
This is the heart of the pipeline. This is where the AI does its second, and more interesting, job.
Each SKU is scored on five dimensions, each dimension computed by a small, specialized model:
Score(S) = w1·V(S) + w2·L(S) + w3·C(S) + w4·M(S) + w5·R(S)Where:
V(S) is the visual consistency score. How well does the product's hero shot and product shots align with the brand's visual language?
L(S) is the linguistic alignment score. Does the product's description, tagline, and review corpus match the brand voice?
C(S) is the commercial consistency score. Does the price tier, channel mix, and inventory depth fit the brand's commercial positioning?
M(S) is the market resonance score. How well does the product perform in the market? Repeat purchase, co-purchase, return rate, review sentiment.
R(S) is the relational coherence score. Does this product sit well within the brand's product family? Does it complement or conflict with siblings?
Each sub-score is between 0 and 1. The weights $w_1$ through $w_5$ are not fixed. They are learned per brand from a small labeled set of past campaign decisions. The brand's historical choices — which products were chosen for campaigns, which were not — become the training signal. We use a small, efficient ranker: a gradient-boosted tree over the five sub-scores, trained on 200 to 500 past decisions per brand.
The ranker is fast. It scores 2,400 SKUs in under two minutes.
The explainability step is the part that makes the shortlist useful. For each SKU in the top 50, the pipeline produces:
A one-line rationale. "Strong visual consistency and high repeat purchase, but price tier is above the brand's premium anchor."
Three to five evidence pointers. "See hero shot #412, brand voice note #88, P&L row #2041, co-purchase graph edge #1192, review corpus #307."
A small bar chart of the five sub-scores, so a human can see where the score comes from.
The human reader is not asked to trust the model. They are shown the evidence. They can check the hero shot, read the brand voice note, look at the P&L row, and decide if the model is right. If it is wrong, they can adjust the weights and re-run. The pipeline is interactive.
This stage takes about nine minutes end to end, including the explainability generation.
Stage 4: Package and deliver
The last stage is the one that makes the artifact usable. The pipeline produces:
A ranked shortlist of 20 to 50 SKUs, each with a score, rationale, and evidence pointers.
A comparative view. A bar chart of the five sub-scores for the top 20, so the reader can see which dimensions drive the ranking.
A gap analysis. Which SKUs are close to the cut? Which dimensions are weak across the shortlist? If the top 20 are all strong visually but weak commercially, the reader knows that is the lever to pull.
A provenance map. For each SKU, a small graph showing which raw assets contributed to its score.
The artifact is delivered in three formats: a structured JSON for downstream systems, a plain markdown document for human reading, and an interactive HTML page for the marketing team.
This stage takes about two minutes.
Why 30 minutes, not 30 hours
The time budget is the point. A 30-minute pipeline means you can run it before a campaign meeting. You can re-run it when a new product is added. You can A/B test two different weight sets. You can run it for two different brands and compare.
A 30-hour pipeline is a project. A 30-minute pipeline is a tool.
The 30-minute budget is not a constraint imposed by the user. It is a constraint imposed by the pipeline designers. Every stage has a time budget, and if a stage blows its budget, the pipeline degrades gracefully. If the vision-language model is slow, the pipeline falls back to a lighter visual model. If the ranker is slow, the pipeline falls back to a simple weighted sum. The artifact is always produced. It may be slightly less precise, but it is always on time.
What the pipeline does not do
It is worth being honest about the boundaries. The pipeline does not:
Decide which products to launch. It ranks candidates. A human makes the call.
Generate new product concepts. It evaluates existing ones.
Replace the brand team. It gives the brand team a faster, more consistent view of their own assets.
Guarantee that the top-ranked SKU is the best one. It gives you a defensible, explainable, evidence-backed ranking.
The pipeline is a tool for human decision-making. It does not replace judgment. It compresses the information the judgment is made on.
A small example
Say you run the pipeline on a mid-size consumer brand with 1,200 SKUs. The output looks like this:
Rank SKU Score Rationale
1 SKU-412 0.94 Strong visual and linguistic alignment; high repeat purchase
2 SKU-887 0.91 Excellent market resonance; price tier slightly above anchor
3 SKU-204 0.89 Coherent with product family; strong review sentiment
4 SKU-561 0.87 Visually consistent; moderate commercial fit
5 SKU-330 0.85 Linguistically aligned; strong co-purchase network
...Each row has a bar chart of the five sub-scores. Each row has three to five evidence pointers. The marketing team opens the HTML page, scrolls through the top 20, checks the evidence, and walks into the campaign meeting with a shortlist they can defend.
Total time from "run the pipeline" to "I'm in the meeting with a shortlist": 24 minutes.
The design principle
The pipeline is built on one principle: keep the modalities separate until the scoring stage, and keep the scoring stage explainable.
Collapse modalities too early and you lose the ability to reason about which dimension is driving the score. Collapse the scoring stage into a black box and you lose the ability to trust it. Keep both open, and you have a pipeline that is fast, precise, and defensible.
That is what a 30-minute pipeline is. Not a trick. A small, well-ordered stack of components doing exactly the job they are good at.
