Conix

Context

Conix was an applied-AI platform for the early stages of architectural design. I co-founded it and ran product. The bet was straightforward: the first two weeks of any architectural project are spent producing massing studies, program allocations, and feasibility models — work that a generative system could now accelerate enormously if the interaction was designed right.

We started the company in 2021, well before ChatGPT, when “generative AI” in this context meant diffusion models and early transformers applied to geometry and spatial programming.

Research & discovery

We embedded with four architecture practices — one global, three regional — for the better part of a quarter. The headline finding: the bottleneck in early-stage design isn’t producing options. Architects can sketch options all day. The bottleneck is evaluating them against the dozens of constraints a project actually has — zoning, daylighting, floor-area ratios, the client’s unspoken taste.

That meant a generative product couldn’t succeed by generating more. It had to generate and evaluate, and it had to do the evaluation in terms the architect already used. A tool that produced a thousand beautiful massings but didn’t tell you which three satisfied the setback rules was worse than useless — it was noise.

We also learned that architects are visual reasoners who still argue verbally. Any interaction that forced them into either pure form-filling or pure prompt-typing felt foreign. They wanted to sketch a shape, describe a constraint, and have the system do both at once.

Approach

The product took the shape of a canvas with an embedded reasoning layer. The architect worked in the medium they already knew — spatial arrangement, massing, program diagrams — and the system ran evaluation continuously in the background, surfacing violations and suggestions without hijacking the workspace. We treated AI as an always-on second opinion, not as the driver.

As CPO I made the call early that we would ship one narrow, deeply-done workflow rather than a generalist tool. That shaped everything — hiring, data collection, fundraising narrative.

Solution

The main canvas — a mixed-modal workspace where spatial edits and text constraints compose into a live feasibility model.

Massing exploration, with generated alternatives ranked by how they score against the project’s active constraints.

The constraints surface — zoning, program, budget — expressed as editable objects the architect can loosen, tighten, or reprioritize on the fly.

The generation flow — structured so the architect is always in the loop, never waiting for a black box.

Evaluation against code and programmatic rules, with plain-language explanations of what each flag means.

The project dashboard — showing where each feasibility study stood, what was blocking, and what needed a human decision next.

Collaboration — the place where a lead architect reviews and comments on the work a junior has framed up, without either of them leaving the canvas.

Outcome

We raised $1.3M in pre-seed capital from firms focused on vertical AI and the built environment. Shipped to paying architecture practices. The company ran for two years before I moved to the Scale AI / government AI work that followed.

Reflection

Two things from Conix shape how I think about applied AI now. First, interaction design decides whether a model-centric product earns trust — and trust decides retention. Second, research comes before model work, not the other way around. The best prompt in the world can’t fix a workflow that misunderstands the user’s actual decision.