Skip to content
Brian Bautista, home
Back to projects

StockPull

One search box over the stock libraries a photo editor lives in, with every result ranked and every rank showing its working.

Client
Trusted Media Brands
Year
2026
Status
Live
01

The hour before the judgment

Picking the picture for an article took thirty to sixty minutes. Getty in one tab, Adobe Stock in another, Woodwing in a third, the brief on one screen and the brand's unwritten rules on a sticky note. Five to ten searches, run one at a time, and the results reconciled by eye. None of that is the editorial judgment — it is the hour of tab-work in front of it.

  • 01Library AOwn login
    Own keyword box
  • 02Library BOwn login
    Own keyword box
  • 03Library COwn login
    Own keyword box
  • 04Library DOwn login
    Own keyword box
02

Why believe the order

Tags find what a picture was labelled. That is half the score. The rest is how much the library is trusted for this brand, and whether the picture looks like the thing described. Every card shows its own arithmetic.

Keyword search

Matches the words

  1. The query
  2. Tags and captions
  3. Whatever someone typed

So you search again. And again. Then you give up and scroll.

CLIP visual scoring

Matches the picture

  1. The query
  2. A visual embedding
  3. The images that look like it

One pass. Nothing to re-run, nothing to compare by eye.

StockPull runs inside TMB, so the real interface is not mine to publish. Every screen on this page is redrawn from the tool it depicts.

Concept mockup of a visual-similarity view: a large reference photograph of a fog-covered pine ridge on the left, and on the right a "Visually similar" row of four closely matching misty forest images, each above an orange relevance bar that steps shorter from left to right, under the caption "Ranked by visual similarity".
Concept mockup. One reference photo, and the four the system scored as its nearest visual neighbours.
03

One box

Paste the article, pick the brand, and about sixty seconds later there is a ranked grid. It reads the brief, plans five to ten search angles, runs them in parallel, drops anything the brand's rules forbid, ranks what is left, and strips the near-duplicates. Every card carries a Match % you can open: half of it is how well the caption fits the brief, a third is how much that library is trusted for that brand, and the rest is whether the picture actually looks like the thing described. The mix is tunable per brand. The editor still picks — the tool does the hour in front of the picking.

Concept mockup of a dark image-search interface: a sidebar reading StockPull above Search, Collections and History, one wide "Search for images…" field, a row of filter chips for Photos, Orientation, Color and Aspect Ratio, and a grid of unlabelled landscape and architecture photographs. One tile is outlined in vivid orange with a partly filled orange relevance bar beneath it.
Concept mockup. One grid, one order. The outlined tile is the top match and the bar beneath it is its score.
04

Behind the one box

Four heterogeneous library APIs behind one gate, and a scoring pass on the way back. The diagram is the entire system. There is no sixth stage hiding behind it.

  1. 01 Query

    • Photo researcherTypes the query once(Human)

  2. 02 Gate

    • Identity-Aware ProxyOne login for all four(Process)

  3. 03 Fan-out

    • Library A(Source)
    • Library B(Source)
    • Library C(Source)
    • Library D(Source)

  4. 04 Score

    • CLIP visual scoring(Model)

  5. 05 Result

    • One ranked list(Output)

Four vendor libraries, queried in parallel behind one login. They are lettered rather than named because they are TMB's suppliers, not part of what I built.

Concept mockup of a "Querying sources" panel: four stacked source-library rows, each with a partly filled progress track and a gray "Querying" status, gathered by an orange connector into a single merged "Ranked results" row whose count slot shows an em dash instead of a number.
Concept mockup. The panel mid-query: four progress tracks, and a result count that has not resolved yet.
Runs on
Google Cloud Platform
Live for
Six TMB brands — Birds & Blooms, Family Handyman, Reader's Digest, Taste of Home, The Healthy, TMB Global
Sources
Getty, in production. Adobe Stock, Woodwing and OrangeLogic are built and waiting on access, not on engineering.
Used by
Photo researchers and editors
Link
None. It sits behind the corporate login.

Search the idea, not the keyword.

Outcome

If your team queries several vendor APIs by hand and lines the answers up by eye, this is the shape of the fix: one gate, one fan-out, one ranking layer.

TMB brands live in production
6TMB brands live in production
Saved per article — team estimate, not yet measured
~25 minSaved per article — team estimate, not yet measured

Picking the image used to take thirty to sixty minutes an article: five to ten searches, run by hand, across libraries that don't talk to each other, checked against brand rules that live in somebody's head. With StockPull it takes six to eleven. The editor still chooses — the tool does the part before the choosing. Those timings are the team's own estimate, taken from interviews and not yet measured in production, and the deck they come from says so on the slide. The number that is not an estimate is six: the brands running it today.

05

The credits

Every part of it is mine: the search UX, the ranking layer and its scoring mix, the sign-on gate, and the federated search layer — four library integrations, of which Getty is live in production and Adobe Stock, Woodwing and OrangeLogic are built and waiting on access rather than on engineering. Creative Technologist, Trusted Media Brands.

  • Google Cloud Platform
  • Identity-Aware Proxy
  • CLIP visual embeddings

Start a project

What do you need built?

Tell me what the thing has to do and who has to use it after I leave. We'll get on a call, and I'll tell you what it takes.

Start a project
Next

Analytics Dashboard

Trusted Media Brands · 2026

Back to projects