← BACK

How do you give an expert enough control to trust an automated system?

For construction engineers, errors in design and coordination are a primary source of risk, rework, and waste. This project is about a generative AI tool, a plugin in Revit, that automates sustainable building design. The main challenge was balancing user control with automation and making the experience as seamless as possible.

COMPANY
Augmenta AI
TIMELINE
4 months
TEAM
CEO · Design Lead · Engineers
TOOLS
Figma · Miro · Revit
Augmenta AI — generative conduit layout tool in Revit

MY ROLEOwned constraints + options comparison features. Research through to final designs.

THE WORLD THEY WORK IN

Errors in construction compound fast

Conduit misrouted
AI ignores no-go zones
🔧
On-site rework
Crew time lost
📈
Cost overrun
Project budget hit

A misrouted conduit means rework on-site. Rework means cost overruns. On a building project, that's not a UX problem. It's a business failure.

Engineers already knew how to do this job. What they needed wasn't a system that replaced their expertise. They needed one that handled the repetitive, error-prone parts without getting in the way of their judgment.

TWO BLOCKERS

Why engineers weren't adopting

01
Up to 20% of AEC time lost to rework

The AI ignored what engineers actually know

Fully automated layouts sound efficient. Until the output routes conduit through a structural blockout, or places a panel in a no-go zone the contractor flagged weeks ago. The AI was fast. It was also wrong in ways that took longer to fix than doing it manually. Engineers weren't slow to adopt. They were right to be skeptical.

02
Up to 30% of materials wasted from rework

Comparing options was a separate job

Even when layouts were usable, engineers had no way to evaluate them side by side. Cost vs. construction time vs. maintainability: these tradeoffs exist on every project, and they were invisible inside the tool. Exploration happened offline, in spreadsheets, after the fact.

DESIGN CHALLENGE

Making automation useful for someone who knows more than the system

The answer wasn't to make the AI smarter. It was to change when the engineer was in control. Constraints had to come before generation, not as a workaround, but as the actual workflow. If an engineer could define the rules upfront, the AI's output would be worth looking at.

BEFORE
generate_layout()
→ conduit collision
→ no-go zone violated
→ rework required
AFTER
define_constraints({
blockouts: [...],
no_go: [zone_4, zone_7]
})
generate_layout()
→ 3 valid layouts ✓
MY INVOLVEMENT

Research, build,
get feedback, repeat

I owned the constraint-setting and options comparison features, from first conversation with engineers through to final designs. The work also included creating custom illustrations and annotations directly on 3D diagrams. Engineers think spatially. The interface had to speak that language.

STEP 01
Talk to construction engineers
Conversations about how they actually work: watching their screens, understanding their mental models of conduit routing and spatial constraints.
STEP 02
Build quickly
Rough designs back to the team fast. Not polished, just testable. Getting something in front of people before investing in the wrong direction.
STEP 03
Design review
Internal review with the design lead and CEO. Pressure-testing decisions before taking them back to engineers.
STEP 04
Back to engineers for feedback
Did the constraint model match how they actually think? Was the annotation legible on a 3D diagram? Second round of feedback before anything was finalised.
WHAT WE BUILT

Three features. One principle.

FEATURE 01

Set the rules before the AI generates anything

Users define device locations, panel positions, blockouts, and no-go zones before running the tool. Routing happens inside a set of rules the engineer already trusts, because they wrote them. Less fixing after the fact. More confidence in the output from the start.

The annotation work here was critical: engineers think spatially, so I built custom callout labels directly on 3D diagrams.

Constraint-setting interface — engineers define device locations, blockouts and no-go zones before generation
Side-by-side layout comparison — cost, construction time and maintainability visible at once
FEATURE 02

Compare options side by side, not in a spreadsheet

The system generates multiple fully-detailed layouts in parallel. Engineers evaluate cost, construction time, and maintainability together, not sequentially, not offline. The tradeoff is visible. The decision is theirs.

FEATURE 03

Review, edit, and hand off without leaving the workflow

Once a design is selected, it goes straight to Revit. No export dance, no format conversion. The output fits the workflows engineers already use for placement, detailing, and spooling.

Handoff to Revit — selected layout exported directly into the engineer's existing workflow
WHAT CHANGED
Days to hours

Generating layout options went from days to hours, because constraint definition made the AI output worth using on first attempt.

Rework down

Constraint violations dropped. When engineers set the rules before generation, the AI stopped routing conduit into places it should not go.

Trust built

The shift happened when the interface started speaking the engineers’ spatial language: constraints annotated in context, decisions visible, tradeoffs surfaced.

LOOKING BACK

The messy part of this project wasn't the UI. It was figuring out where human judgment ended and where automation could begin, and making sure that handoff was legible to someone who'd spent 20 years doing this by hand.

That's still the problem I find most interesting. Not how to make AI more powerful, but how to make the point of contact between expert knowledge and machine output feel honest and trustworthy.

Want to hear the
real story?

The prototype is one version of this. I'd rather walk you through the decisions, the wrong turns included.

Get in touch →