Case Study
At SWAPP.AI, I worked on a domain-specific AI assistant designed for architects - a conversational system capable of interpreting architectural prompts and translating them into precise changes across building plans.
The assistant allowed architects to communicate naturally through written prompts or sketches and receive real-time feedback, from identifying compliance issues to adjusting walls, generating annotations, and applying related updates across similar elements.
Our internal architectural team served as both expert users and active contributors. Their daily workflows, structured feedback, and ongoing validation helped us refine how the system understood architectural intent and responded within the constraints of building codes, construction standards, and firm-specific guidelines.
The goal was to reduce repetitive manual work, surface code violations earlier, and help architectural teams iterate faster without compromising accuracy or design control.
The challenge was not simply to automate architectural tasks - it was to make complex AI behavior understandable, controllable, and trustworthy.
Designing around evolving inputs and real architectural behavior
This feature was developed in close alignment with the architectural team, who served as both expert users and internal testers. Their real-world workflows shaped how the system interpreted prompts and generated visual changes across architectural plans.
Instead of following a fixed workflow, we worked iteratively - responding to the changing needs of our internal users, expanding functionality based on real usage, and refining the system's behavior through hands-on feedback.
The product required constant calibration between UI, logic, and architectural accuracy. Every interaction helped define how the assistant understood architectural intent and translated it into precise, standards-compliant updates.
Supporting creative workflows within technical constraints
Architects spend hours working on complex floor plans that must comply with strict building codes. One of the biggest challenges they face is identifying and fixing compliance issues while maintaining their design intent.
Before this feature, much of that work was done manually - verifying corridor widths, wall placements, or annotation accuracy required constant zooming, checking, and redrawing.
Users needed a way to:
The challenge was to build a tool that supported architectural precision while remaining as intuitive and conversational as possible. We had to balance flexibility and structure - allowing for creativity without compromising code compliance.
Translating research into a structured AI workflow
Before designing interfaces, we mapped how architects, AI, and building regulations would interact throughout the review process. The goal was to understand not only what users needed to accomplish, but also how information should flow between human decisions and AI-generated suggestions.
Instead of designing isolated screens, we first designed the system itself - defining how prompts would be interpreted, how validation would occur, where user confirmation was required, and how architectural knowledge would continuously improve the assistant's responses.
This early system mapping helped define:
By solving the workflow first, the interface became a direct expression of the system's logic rather than a collection of disconnected screens.
How architects collaborate with AI through a guided annotation workflow
Instead of asking users to learn a new system, the experience was designed around a familiar architectural workflow.
Architects simply describe the changes they want to make in natural language while working directly on top of the floor plan.
The assistant analyzes the request, identifies the relevant locations, highlights potential issues, explains its reasoning, and gradually builds a structured understanding before suggesting architectural updates.
Throughout the interaction, the architect always remains in control, reviewing every recommendation before any change is applied.
The conversation experience focused on four principles:
The interface became less about operating software and more about creating a reliable dialogue between architectural expertise and artificial intelligence.
01
The conversation starts with a simple natural-language request while the architectural drawing is loaded into the workspace.
As the assistant begins processing the prompt, the system communicates that analysis is currently in progress, providing immediate feedback instead of leaving the user waiting without context.
This small interaction helps establish trust by making the AI's thinking visible from the very beginning.
02
Instead of making silent changes, the AI highlights the affected area directly on the drawing, explains why it requires attention, and provides architectural context.
Supporting information and quick actions appear alongside the annotation, allowing architects to understand the recommendation before deciding how to proceed.
This transparent reasoning transforms AI from an automated tool into an explainable design partner.
03
Instead of automatically modifying the project, each suggestion is presented as a reviewable action.
Architects can evaluate the recommendation, understand its implications, and choose the most appropriate next step before applying any change.
Keeping the human in control ensures confidence while maintaining an efficient collaborative workflow between architect and AI.
Expanding a single design decision across the architectural model
One architectural change rarely exists in isolation.
When a wall is added, removed, or modified, the surrounding environment is often affected as well.
Instead of stopping after a single correction, the assistant analyzes nearby architectural elements and identifies additional locations that should be reviewed.
This allows architects to validate a broader set of recommendations while keeping full control over every decision.
The goal is not to automate design decisions blindly, but to surface connected opportunities that would otherwise require manual investigation.
Making AI recommendations understandable, reviewable, and reliable
The success of this product depended on more than accurate AI predictions. It depended on helping architects understand what the system was doing, why it reached each recommendation, and when human judgment was still required.
Instead of hiding complexity behind automation, every interaction was designed to make the assistant's reasoning visible. Recommendations were introduced progressively, supported with contextual explanations, and always presented as reviewable actions rather than automatic decisions.
This approach encouraged confidence without removing control. Architects could move faster while remaining responsible for every meaningful design decision.
Designing explainable interactions proved just as important as designing intelligent ones. Trust was built through clarity, transparency, and predictable system behavior, not through automation alone.
Designing transparency without sacrificing efficiency
Throughout the project, one principle consistently guided every design decision: architects should always understand what the AI is doing and why.
Instead of hiding complex logic behind automation, the interface exposes the assistant's reasoning through progressive feedback, contextual explanations, and reviewable recommendations.
Every interaction was designed to balance speed with confidence, enabling architects to work faster without losing visibility or control. The assistant analyzes architectural intent, detects related opportunities, and surfaces meaningful recommendations, while leaving every important decision in the architect's hands.
This approach transformed AI from an unpredictable automation tool into a collaborative design partner, one that supports professional judgment instead of replacing it.
The result is a collaborative experience that reduces repetitive work, increases confidence in AI-assisted workflows, and allows architects to focus on higher-value design decisions.