OpSpring.ai
All case studiesDesign & Construction

A Boutique Interior Design Studio

Sixteen hours per project spent retyping specifications the studio had already written once. We built the design board that captures them the first time.

16 hrs
retyping per project, studio estimate
468
product links in one deck
6 of 821
spec notes machine-readable
114
vendors, top 10 are half

The challenge

The studio designs residential interiors in presentation decks, then re-enters everything in those decks into a project management tool to produce line items and purchase orders. They estimated the re-entry at sixteen hours per project and wanted to know whether it could be automated.

Rather than take the estimate on faith, we asked for one real project file and took it apart with a script before scoping anything. The file was one gigabyte: 2,204 individual files inside a single document, 68 slides, roughly 25 room sections, 468 unique product hyperlinks across 724 link instances pointing at 114 vendor domains, and 821 free-text specification notes, about twelve per slide.

Sixteen hours across 468 items is 2.05 minutes per item. That is what the clip-and-paste workflow costs: paste the link, wait, highlight the fields, add the size and finish, save, next. Their estimate was arithmetic, and it was correct.

The finding that reframed the engagement: of those 821 specification notes, six contained a dimension a computer could read. Six notes, not six percent.

The notes themselves are complete. "Artistic tile ocean blue vein cut, wainscot halls in entire room, Schluter greige, Nordic Timber 48 tall, 8 x 48 tiles, run one piece vertically, Tec sterling power grout." Everything a person needs to place that order is there. It is sitting in a text box, beside an image, near a hyperlink, with nothing connecting the three.

So the sixteen hours was a data recovery problem more than a data entry problem. Every fact needed to build those line items was typed by a person once, during design, then flattened into a slide where the image, the link and the note stopped being related. The team was going back into their own deck to reassemble information they already had.

The solution

The first plan was to read finished decks and pull the product data back out. The teardown argued for something simpler: if the facts come apart the moment a product lands on a slide, stop putting them on a slide.

So we built Bench, a design board that replaces the deck. A designer drops a product image onto the board, removes the background in one click, and arranges it in a room the way they did on a slide. Behind every image is a product record: link, vendor, price, SKU, finish, dimensions, notes, room and quantity. Quantity is separate from what is shown, because a room may show one of four chairs.

The designer does not type that record from scratch. The Bench Clipper, a browser extension, reads the product page inside the designer's own browser, so it gets past vendor sites that block automated tools and sees trade pricing when the designer is signed in to the vendor. Anything it cannot find stays blank for the designer to fill in.

Each room has a budget that adds up as products are placed. The client gets a share link and approves or declines each item; designer notes, cost and markup never appear on it. The export goes into the studio's project management tool on that tool's own import template, one file per room, with each product's photo alongside.

Boards are saved on the server, so everyone in the studio opens the same project from any computer. A save from an out-of-date copy asks before it overwrites anything.

AI architecture

The clipper reads a product page in two passes. The first takes the structured product data many vendor pages already publish, and costs nothing. The second sends the page's visible text to a small language model for what the structured data misses: name, retail and trade price, SKU, finish, materials, size and lead time.

Any price, SKU or size the model returns that does not appear on the page is dropped. A blank field prompts the designer to look. An invented one turns into a wrong order.

Background removal runs through a hosted service, because the in-browser model could not run on the studio's Macs. When a photo is a full room shot rather than a product on a plain background, Bench declines with a reason instead of guessing at the edges.

The results

The studio uses Bench on real projects.

Each product is entered once, on the board where the room is designed, and everything after that reads from the same record: the room budget, the client's approval page and the import file for the project management tool. Nobody goes back into a finished deck to reassemble what they already knew.

Pricing is on the board during design. Room totals add up as products are placed, instead of waiting for an order list to be built after the design is done.

Bench follows the designers' own description of their process: drag the image in, link it, remove the background, add a note, arrange it in the room, label the room. It does those steps and leaves out the rest of the presentation software they were not using.

The order of the plan changed along the way. The original proposal put the design board last, after extraction and integration were proven. Building it first meant there was nothing left to extract: the data is captured when the product is placed, so it never has to be recovered.

Engineering notes

The board looks simple on purpose. These are the decisions underneath it.

01
A product record that has a picture
A board item is a product record with an image, not an image with notes attached. Position, size, crop and layer order are kept apart from the record, and the record's fields are named to match the project management tool's import template, so the export is a mapping rather than a translation.
02
Quantity is not what is shown
Every image of the same product shares one record, so showing one chair and ordering four is ordinary. A copy dragged into a different room becomes its own record, so an alternate version of a room keeps its own budget.
03
Budgets use the client's price
Room totals use cost plus markup when a cost exists, the way the invoice will price it, and the list price otherwise. A price read from a public web page only ever fills the list price, never the cost.
04
The client sees only what is theirs
The approval link lets the client approve or decline each item. Designer notes, product links, cost and markup never reach it. Each approve or decline is saved as its own entry, so a client clicking quickly cannot lose a decision.
05
No silent overwrites
Every save carries a version. A computer holding an out-of-date copy is told the board changed elsewhere and chooses to load the newer one or keep its own. Nobody loses a morning of work to someone else saving.
06
Export by room, photos included
The import template has no column for a room or an image. Bench writes one file per room, split at the 500-row import limit, with a folder of photos beside each file, numbered in row order and named the way each item will appear after import.
Known limits

Several people can open the same board, but it is not live co-editing: other computers’ saves arrive every 15 seconds. Vendor pages behind a trade login only give up trade prices when the designer is signed in to that vendor. The import cannot place items in rooms, so the team moves each imported file into its room, one batch per room.

The information needs to stop getting thrown away. Nobody needs to type faster or work smarter.
From the engagement notes·On why the 16 hours is not a data entry problem

Tech stack

  • Design canvas
  • Browser extension
  • Language model extraction
  • Hosted background removal
  • Postgres
  • Import file generation

Ready to streamline your operations?

Schedule a 15-minute diagnosis call to see if AI can solve your biggest bottlenecks.