AI search & internal toolsIn production

Past creative, found with one plain-English search.

Designers search the agency's archive by idea, angle or format and jump to the Figma source, instead of an afternoon of digging through shared drives.

129client folders migrated
21,065designs in the library
Design library gallery showing a plain-English search and filters for angle, funnel stage and asset type
Rebuilt interface · illustrative data

The business problem

Past creative lived across 129 client folders in a shared drive, behind a filename convention that was not always followed. Designers could browse files, but they could not search by idea, messaging angle or visual concept. Useful work was hard to find, so it was rarely reused.

What changed

Yash built a resumable migration pipeline that copies files from the drive into cloud storage and records each design in a database. A five-step AI tagging run then adds visual elements, the messaging angle, the copy in the image, a plain description and a check on whether each file is really an ad.

The gallery search turns a plain-English query into an embedding and ranks designs by meaning, with structured filters on top. A Search in Figma button takes designers to the source file. The same search is also open to AI assistants through a connector with seven tools.

Before & after

AreaBeforeAfter
Finding referencesBrowse folders and guess filenamesDescribe the idea and apply filters
Asset contextInconsistent namingAngle, visual elements, copy and description on every design
Reusing workFind the file, then hunt for its sourceSearch in Figma from the design itself
Starting new workCollect references by handPick references, then generate variants in the builder

The business value

129 client folders became one searchable archive of 21,065 designs. Tagging, review and sync run as separate steps, so the catalogue stays maintainable as new work arrives from the shared drive.

How to read the return

The outcome is a searchable, reusable archive. Search time and reuse rates were not measured. The design count is from the project's own notes (October 2026).

Why the system is dependable

The filename parser takes what it can and flags uncertainty instead of guessing. Designers can edit tags by hand, and those edits are marked manual so the AI does not overwrite them. Embeddings carry a content hash, so unchanged designs are skipped when the library is re-embedded. A per-client consent flag controls which work may appear on public showcase pages, and new clients start private.

A closer look at the build

Design detail panel with AI tags, ad copy, alt text and a Search in Figma buttonEvery design: tags, copy and a route to its Figma source (full size, opens in a new tab)Builder form with library references and four generated variantsGenerate new variants from library references (full size, opens in a new tab)

Interface mock-ups use fictional brands and ad copy. The real library holds client work and is internal only.

Could this work for your team?

If your best previous work is buried in folders, semantic search can turn that archive into an everyday working tool.

Talk through a similar project
Keep exploring

More case studies

All case studies