ENTERPRISE SOFTWARE
KPS Mart
Cutting Time-to-Proposal Without Adding Headcount
How AE Studio built a computer vision and LLM workflow that transforms KPS Mart's furniture proposal process from weeks of manual work into a 22-minute automated draft, letting designers focus on the final 20% instead of the first 80%.
"It took me 22 minutes from when I started until it was done, and it used to take that project team a couple of weeks."
Viktor, Senior Director at KPS
THE CHALLENGE
The problem.
KPS Mart has been designing interior spaces and selecting furniture for clients for over 30 years, primarily for offices and gyms. Their core deliverable is a general arrangement: a detailed design proposal showing furniture placement, product selections, and cost estimates tailored to the client's space and budget.
The problem was time. Creating a single proposal required designers to manually interpret CAD floor plans, identify room types and zones, cross-reference a product library of over 10,000 items across multiple manufacturers, and compile SKU-level selections with budget options. A project could take a team weeks from first floor plan to finished proposal.
This bottleneck capped how many concepts KPS could present to clients and how quickly they could iterate when clients asked for changes. The business wanted to increase the number of proposals generated, reduce the cost of creating each one, and give clients more options faster. But hiring more designers was not a scalable answer. KPS needed a way to automate the heavy lifting of proposal generation while keeping their designers in control of the final output.
THE SOLUTION
What we built.
Computer Vision for Floor Plan Interpretation
The first step in any KPS proposal is understanding the space. Designers work from CAD files exported as PDFs, which contain floor plan layouts, architectural symbols, room labels, dimensions, and zone annotations.
AE built a computer vision pipeline that ingests these CAD and PDF floor plans and automatically identifies furniture placement areas, architectural symbols, room types, zone boundaries, and spatial dimensions. What previously required a designer to manually interpret is now extracted programmatically, creating a structured representation of the space that feeds directly into the recommendation engine.
Object Classification and Furniture Identification
Beyond reading the floor plan layout, the system classifies each furniture element in the plan, identifying item types and assigning accurate counts per zone. This object classification step ensures the recommendation engine knows exactly what categories of furniture are needed in each area of the space before making any product selections.
This is a critical accuracy gate. Misclassifying a lounge chair as a desk chair, or undercounting workstations in an open plan area, would cascade into incorrect proposals. The classification layer was built to handle the variability in how different architects and designers represent furniture in CAD exports.
SKU Recommendation with Budget Tiering
With the space understood and furniture needs classified, the system matches requirements against KPS's product library to generate selections. KPS manages over 10,000 products across multiple manufacturers, each with its own dimensions, finishes, and price points.
The recommendation engine accounts for spatial constraints, client requirements, and budget to produce three tiers of options for each proposal: good, better, and best. This gives clients real choices and gives KPS designers a structured starting point rather than a blank page. Previously, generating these tiered options manually was one of the most time-consuming parts of the process.
Product Catalogue Digitization
Before the recommendation engine could work effectively, KPS needed their product library in a structured, machine-readable format. Manufacturer data arrives in inconsistent formats: PDF data sheets, product images with embedded dimension callouts, and multi-language specification tables.
AE built a computer vision pipeline to extract furniture dimensions, finishes, SKU codes, and pricing from manufacturer images and data sheets into structured JSON. This digitization work transformed KPS's 10,000-product catalogue into a format the recommendation engine could query reliably, and established a repeatable process for ingesting new products as the catalogue grows.
Human-in-the-Loop Designer Dashboard
Automation handles the first 80% of the proposal. The final 20% stays with KPS designers. AE built a review dashboard where designers can inspect the AI-generated proposal, flag any misclassified furniture items, swap product selections, and approve the final picks before they flow into the quoting system.
Critically, designer adjustments in the dashboard do not alter client-facing deliverables directly. The workflow keeps the AI output and the final proposal cleanly separated, ensuring designers remain in control of what goes to clients while still benefiting from the speed of automated generation.
HOW IT WORKS
The details.
Reading Floor Plans the Way a Designer Would
The first step in every KPS proposal is understanding the space. Designers work from PDF exports of CAD files that contain room layouts, symbols, dimensions, and zone labels. We built a pipeline that reads these files automatically, identifies furniture placement areas, room types, and spatial boundaries, and creates a structured representation of the space. What previously required a designer to interpret manually is now extracted in seconds.
Counting and Classifying Every Piece of Furniture
Beyond reading the layout, the system identifies each furniture element in the plan, classifies what type it is, and counts how many are needed in each area. This is a critical accuracy step. Misclassifying a lounge chair as a desk chair would cascade into a wrong proposal. The system was built to handle the variability in how different architects and designers represent furniture in their drawings.
Three Price Tiers, Generated Automatically
Once the system knows what furniture is needed and where, it selects products from KPS's catalogue of over 10,000 items across multiple manufacturers. It accounts for spatial constraints, client requirements, and budget to produce three tiers of options for each proposal: good, better, and best. Previously, generating these tiered options manually was one of the most time-consuming parts of the process.
Turning a Paper Catalogue Into a Machine-Readable Database
Before the recommendation engine could work, KPS needed their product library in a format the system could use. Manufacturer data arrives in inconsistent formats: PDF data sheets, product images with dimension callouts, and multi-language specification tables. We built a pipeline that extracts dimensions, finishes, SKU codes, and pricing from these materials into structured data. This process also works for new products as the catalogue grows.
Designers Review and Approve, Then It Goes to the Client
The AI handles the first 80% of the proposal. The final 20% stays with KPS designers. We built a dashboard where designers can inspect the AI-generated output, fix any misclassified items, swap product selections, and approve the final picks. Designer changes in the dashboard do not alter client-facing deliverables directly. The workflow keeps the AI output and the final proposal cleanly separated so designers stay in control of what the client sees.
OUTCOMES
What shipped.
Time-to-proposal reduced from weeks to 22 minutes
Designers now handle only the final 20% of the proposal process
10,000+ product catalogue digitized and made machine-queryable
Good/better/best budget tiering generated automatically per proposal
Increased proposal capacity without adding headcount
KEY TAKEAWAYS
What we learned.
- Computer vision unlocks structured data from unstructured design files. CAD exports and PDFs contain rich spatial information that, once extracted automatically, can drive downstream automation that was previously impossible at scale.
- Automating the first 80% is more valuable than automating everything. Keeping designers in the loop for final review preserves quality and client trust while eliminating the manual work that actually constrained capacity.
- Product catalogue quality is a prerequisite for recommendation accuracy. Digitizing 10,000 products into structured JSON before building the recommendation engine was unglamorous work but essential to the system performing reliably.
- Budget tiering increases proposal value. Giving clients good/better/best options in every proposal creates more decision-making flexibility and positions KPS as a more thorough partner than competitors who deliver a single option.
- Internal tools can outperform client-facing products in ROI. Starting with an internal designer tool rather than a customer self-service product allowed KPS to move faster, maintain quality control, and prove the concept before broader deployment.
IN SUMMARY
Bottom line.
In summary, KPS Mart's proposal process went from a weeks-long manual effort to a 22-minute automated workflow, without adding headcount or sacrificing design quality. Furthermore, As a result, By combining computer vision for floor plan interpretation, an LLM-powered recommendation engine, and a human-in-the-loop review dashboard, AE Studio gave KPS's designers a system that handles the heavy lifting and leaves them to focus on what matters most: delivering great spaces for their clients.
FAQ
Frequently asked.
What problem was KPS Mart trying to solve?
How does the floor plan interpretation work?
How does the system handle KPS's large product catalogue?
What does the good/better/best tiering mean in practice?
What role do designers play after automation?
How much time does the system save?
LET'S TALK
Bring us the hard problem.
We'll bring the team that ships.