Our Commitment to Architects & Co-Creation

At TestFit, our greatest passion is the meaningful creation of buildings — specifically, the kinds of buildings that typical human beings use on a regular basis. We are driven to improve the quality and quantity of commodity architecture. We are disrupting the building feasibility process in the housing, hotel, office, retail, parking, and multifamily industries, and ultimately aim to assist in ending the worldwide housing crisis.

Some Context

In my capacity as Co-Founder and CEO at TestFit Inc I oversee the creation of algorithms that design buildings. We automate feasibility studies for commodity buildings. We sell our algorithms in the form of a configurator — essentially "excel" for buildings. We package it in software and sell that software to architects, real estate developers, investors, and general contractors. You will notice that 2/3 of our users are architects. That's right: we sell practical automation tools to architects.

A Brief and Grossly Over-Simplified History of Automation

Cities only exist because a hunter-gatherer decided they had had enough of hunter-gathering and the modern industry of agriculture was born over 8000 years ago. Those early innovators found that tilling soil was back-breaking work, so they automated it by domesticating large animals that could till soil easily. Fast forward to the 21st century: GPS-enabled combines do all of that work autonomously and architects are still using trace paper.

The digital mechanization of the architecture industry started in the 1980s. It's been happening for 40 years, longer if you include mechanical mechanization (looking at you parallel bars, compasses, protractors, and scales). Given the amazing propensity of human innovation, automation is inevitable. It is also true that machines are really bad at doing anything without human direction or intervention.

So, what is the litmus test for whether or not a particular automation should exist? I credit this line of thinking to Travis Connors (Co-Founder at Building Ventures):

If a machine can do it better than a human, a human will never do it again.

One hilarious example of this is Juicero, which raised $140M from Silicon Valley to make a juice press that was less effective than a human squeezing juice with their hands.

The opposite of this is the prevalence of mapping technologies. When was the last time you actually gave someone directions? Just drop a pin. Humans suck at giving directions and following them. Do you know what follows directions perfectly and executes the same exact way every single time? Algorithms.

Doing the Hard Work of Building the Constraint Engine

In March 2018, Bill Allen of Evolve Lab invited me to write a guest post on their blog. I titled it "Let's Face It. Connecting Nodes Sucks" (republished here as Let's Face It, Connecting Nodes Sucks). It walks through some of the early algorithms we built to create our minimum viable product Residential Engine. We reasoned that in order to have an accurate building to then do optimization on, life safety must be taken into account.

We felt the tool must provide real data that confirms our users' planning assumptions to materially provide real value to them. An allowance needs to be made for setbacks. Unit mix and balcony assumptions cannot be ignored as they affect the mass placement and the parking demand. In the real world, buildings are an amalgamation of thousands of parameters, and most "generative design" tools have tens of parameters, are often site-specific and completely un-reusable. By comparison, our most powerful algorithm in late 2020 has a whopping 150 parameters revealed to the user.

Note who sets those 150 parameters: the user. The machine does not decide the setback, the unit mix or the parking ratio. It decides where everything goes once a person has.

A Quick Anecdote About Fear

At one TestFit presentation in Atlanta in late 2017, an architect suddenly stood up and started muttering that he should learn how to be a poet because everything sacred was going to be destroyed by millennials and their computers. We were new and un-tested. We had no co-creation tools, and freaked out many people in the planning world. TestFit is viewed by many in architecture as a threat. It's not unreasonable — humans always resist unknown change.

We acknowledge this fear. If there is only one voice on "Generative Design" it will assuredly get implemented for the aims of that voice. One could sell an algorithm for $100k to developers to have exclusive access to some kind of black box "optimizer." I've led 865 external client meetings in the last three years, and can guarantee that it is much wiser to build a product that has had thousands of architects, developers, and general contractors use, weigh in on, and provide feedback to. Iron sharpens iron, and our future roadmap doesn't lead to black box. It leads to a very sharp co-creation ecosystem.

The release record backs up the "no co-creation tools" part. The first engine, RE 1.011, shipped in November 2017 with no manual editing. Manual Mode arrived in RE 1.014 in February 2018, in the same release as a feedback form. By that record, the architect in Atlanta was looking at a machine he could feed and then accept or reject, but not touch.

Our Promise to Architects: It's Co-Creation, not Desolation

If we don't continue to introduce manual editing and manipulation tools, we are no better than a black box. If we don't keep architects' needs central to our mission, our building configurator has no meaningful future or broader impact on the industry. The future of TestFit is not replacing architects — it's making them dramatically faster at the parts of the job that don't require genius.

Who Decides What

So what does co-creation mean, concretely? When the machine hands you a plan, who decided what is in it, and who gets to change it? "Co-creation" gets used as a mood. Here is the job split I mean by it. In a 2018 LinkedIn comment I put it this way: computers "are great at doing millions of tiny math problems. And Humans are great at understanding complex problems, simply." By 2022 my shorthand was that commodity buildings "should be generated first then edited for context or other constraints by human users." In practice the work splits four ways:

Step 3 is what separates co-creation from a black box. An optimizer that hands back one answer, or locks you out while it thinks, leaves you steps 1 and 4: feed it, then accept or reject what comes out. In 2024 I complained that the topography optimization tools on the market "took the model and loaded it into the cloud and blocked out the user from making live edits while it optimized." That is a human in the loop only at the doors.

The inside corner is the cleanest example. Wrap a double-loaded building around a turn and the solver will happily hand you an inside corner unit, the hardest unit to design, lease and live in (the whole argument is in Multifamily's Problem Child). The machine doesn't know that's a leasing problem. The architect does. In May 2020 my advice was blunt: "Since TestFit is a Co-Creation tool, simply use manual mode to remove them," and that month's release made cores movable so a core could take the corner instead. The machine proposed, the person who knows the market overruled it, and the machine re-counted.

Did We Keep It?

A promise made in 2020 can be checked against what shipped. TestFit's release log is published on this site as the Release Track Record: about 170 features across 53 major releases, November 2017 to March 2026. Read for editing tools, it says three things.

• Editing came early. The first engine (RE 1.011, November 2017) shipped 28 features, none of them for manual editing. Manual Mode arrived three months later, as noted above. The Unit Editor followed in July 2018, multiple schemes in November 2018, the road editor in January 2020 and movable cores in May 2020, all before this essay.
• Editing kept getting rebuilt. Manual Mode has been reworked in 16 major release cycles: extended to lower-density sites in 2021, taken into 3D for garden and low-density buildings and onto industrial sites in early 2022, and rebuilt to run in real time in TestFit 4.0 (July 2023). Only one feature in the record has been revisited more: Generative Design, at 22. When 4.0 shipped I wrote, "Everything is co-creation. Everything is live and connected."
• The optimizer came last. The feature actually named Generative Design launched in July 2024 (v5.000), more than six years after Manual Mode.

Two caveats. The record is our own release log, compiled from release notes. It counts when a feature shipped and how many releases touched it, not how many architects used it or whether they felt served by it. It shows where the effort went, not that the effort worked.

Where the Promise Needs Qualifying

"The future of TestFit is not replacing architects" is a promise, and promises need fine print. Three qualifications, each from my own later posts.

Editable is not the same as harmless. In a 2023 comment I wrote that "TestFit is more a threat to architects than chat gpt or any of the other AIs written about in the article, and most of our effort is in practical automation with architects as customers." Both halves are true. Co-creation decides who holds the pen. It doesn't give back the drafting hours the solver took over, or the pay that came with them.

Control costs simplicity. In 2021 I admitted that "TestFit was very intuitive for the first two years, but adding our most valuable features like co-creation and the configurator makes it hard to just 'get in and drive.'" Every handle you give the architect is one more thing to learn before the first good scheme.

We drifted, and had to walk it back. Shipping the third major revision of Generative Design in TestFit 5.19 (March 2026), I wrote that version 2 "simply generated everything under the sun with limited user control." The fix was this essay's promise again: any parameter can be pinned, and the system regenerates some 1,500 options the instant a filter changes. The company that wrote this essay still built a version of the black box.

Here's where I land. Co-creation, concretely, is three rights the architect never hands to the machine: set the assumptions, reach into the result and change it while the machine re-solves around the change, and reject the answer. A tool that takes away any one of them is a black box, whatever the marketing calls it. Those rights matter more as the option count grows. Man Vs Machine counted twenty iterations for one building; the current generator produces 1,500 the moment a filter moves. Nobody judges 1,500 plans by looking at each one. The judgment moves upstream into which parameters to pin and which filters to trust, and that is architectural work: knowing which constraint is wrong before the machine draws it twenty, or fifteen hundred, times.