I built Caustic after looking again at the documented focused-sunlight failures at Vdara's pool deck and London's 20 Fenchurch St. A concave glass facade can turn a striking rendering into a concentrated-light problem on the ground below it. By the time that problem is discovered after construction, the geometry has stopped being easy to change.
I did not build this demo to claim that a browser can replace an optical specialist. I built it because I kept returning to an earlier question: why should a concept team wait for a mature package before it asks whether reflected sunlight is converging on an occupied plane? The hard part was making that question concrete enough to act on, while keeping the boundary of the answer visible.
Caustic is a Facade Solar-Convergence Firewall for that moment in the design process. It works on synthetic, illustrative facade concepts. A deterministic ray trace follows reflected sunlight through a defined summer-solstice sweep, and a fixed policy gate returns SCREEN-CLEAR, FLAGGED, NEEDS-REVIEW, or ESCALATE. The agents explain the protocol and check the narrative, but they cannot change the calculation or verdict.
I built Caustic so its reasoning can be inspected rather than delivered as an architectural opinion in polished language. Its intended demo route is https://veriprajna.com/demos/ai-architecture-structural-engineering, which will go live with the demo page. It is pre-screening only, not a stamped analysis. That limitation is part of the design, and it became the central lesson of the build.
The concept image that left out the consequence
I started with the most legible case I could make: Crescent Tower, a synthetic 38-storey concave south-glass hotel concept over a pool deck. The form is intentionally attractive. That is why it is useful. Generative tools produce pixels, not load paths, and a persuasive image can conceal the physical question that still has not been asked.
I spent the first part of the build watching a simple facade sketch turn into a solar-study sheet. The FacadeSpec supplies curvature, orientation, dimensions, glazing reflectance, site latitude, and the occupied target plane. The Ray Tracer then samples the facade, reflects parallel sun rays over 15 positions, intersects those rays with the target plane, and records peak concentration, focal surface temperature, and focal dwell.
I chose a defined summer-solstice sweep because the alternative was an answer that sounded precise without being reproducible. The Screening Agent explains why that protocol applies to the concept, but the test itself is deterministic. The Policy Gate uses fixed thresholds: 1.45 times concentration, 80 degrees Celsius focal temperature, and 30 minutes above threshold. No agent can alter those numbers after seeing the result.
When I ran Crescent Tower, the screen returned FLAGGED: 2.07 times peak concentration, approximately 106 degrees Celsius focal surface temperature, and 6.0 hours per day of focal dwell in the tested sweep. I had seen those values in the output before I laid out the panel, but the visual convergence was the moment that stopped feeling like a dashboard metric. The drawing shows the rays gathering over the pool-deck plane. The numbers beside it explain why that location cannot be treated as an incidental reflection.
The illustrative Crescent Tower run is FLAGGED after the 15-position sweep. The panel records 2.07x peak concentration, about 106 degrees Celsius at the focal surface, and 6.0 hours per day of focal dwell over the pool-deck target plane.
I kept the word “illustrative” attached to this case while writing the interface. Crescent Tower is not a customer project, and those numbers are not measurements of Vdara or 20 Fenchurch St. The documented buildings establish the failure mode. This synthetic concept gives a design team something they can inspect before a form becomes expensive to revisit.
I wanted the tool to show the physical consequence of a curved surface before the rendering made that surface feel settled.
I tried to make the ambiguous case look decisive
I made my most useful mistake when I moved from the concave tower to Helix Pavilion, a synthetic freeform doubly-curved diagrid shell. A dashboard wants symmetry: one concept, one sweep, one focal number, one verdict. I initially wanted that symmetry too, because a tidy output feels like progress when you are trying to make a demo explain itself.
I could not support it. Freeform or faceted geometry without a resolvable single curvature radius does not give this pre-screen a confident focal estimate. The geometry exceeds the simplified model's resolution. Producing a number there would turn a limitation into a claim.
I changed the behavior instead of trying to hide that gap. Helix Pavilion returns ESCALATE, asserts no focal figure, and directs the work to full optical simulation and specialist review. The interface calls out the absence directly. Its “Asserted Figure” reads “None.” Its next action reads “Full Optical Sim.” I think that is a more useful screen than an invented temperature that happens to fit neatly in a metric card.
The synthetic Helix Pavilion does not receive a focal-temperature claim. Caustic records ESCALATE, “Geometry beyond pre-screen optical resolution,” no asserted figure, and full optical simulation as the next action.
I remember the temptation in this part of the build as an interface problem before it was an engineering problem. A blank-looking metric feels unfinished. But the absence is the result. The tool has learned enough from the facade description to know that its own approximation does not justify a focal claim.
That refusal also protects the rest of the workflow. It keeps a concept team from mistaking a preliminary screen for a final answer, and it gives the specialist review a clear reason for being called. A useful engineering tool has to preserve the handoff , especially when its most persuasive output would be a number it cannot defend.
The remediated facade changes the tested physics
I went back to Crescent Tower after that failure and built an illustrative same-program alternative: convex curvature with fritted low-e glazing. I did not use it to suggest that a material toggle automatically makes a facade safe. I used it to test whether the concept loop could display a meaningful physical contrast while the geometry was still negotiable.
The remediated configuration returns SCREEN-CLEAR in the tested sweep at 1.10 times concentration, approximately 74 degrees Celsius, and 0.0 hours per day of focal dwell. SCREEN-CLEAR permits continued concept exploration , nothing more. It is not a permit approval, a certification, or a substitute for the downstream optical and licensed review that a built project requires.
I found the contrast more valuable than the isolated red result. The concave concept concentrates reflected rays toward the occupied plane. The convex, fritted configuration spreads the reflected pattern in the tested setup. Putting the two study sheets beside each other made the design consequence discussable in a way a render alone could not. A team can challenge the inputs, inspect the sweep, change the form, and decide what needs the deeper study.
The illustrative remediated Crescent Tower configuration returns SCREEN-CLEAR for this tested sweep: 1.10x peak concentration, about 74 degrees Celsius focal surface temperature, and 0.0 hours per day of focal dwell. The next step remains proper simulation and licensed review.
I deliberately stopped short of turning that comparison into a savings claim. The demo includes a $3,918,060 retrofit-exposure figure, but it is an illustrative calculation from unrounded area and a stated $35-per-square-foot assumption, not a market price, project estimate, customer outcome, or promise of avoided cost. The engineering case stands without pretending to know the price of a project that does not exist.
A gate should leave a trail behind it
I built the Constructability Report because I did not want the screen to be the only place the decision existed. Design review needs more than a colored verdict. It needs the concept input, the screening protocol, raw result, fixed thresholds, critic note, provenance, and the pre-screening disclaimer to travel together.
I had to resist treating the report as a decorative export. The receipt is part of the engineering behavior. A reviewer should be able to see the FacadeSpec that was screened, the defined solar sweep, the numbers that crossed the gate, and the caveat that accompanies the conclusion. The HTML and JSON exports make the screen a record of a constrained decision rather than a transient presentation.
I also put the evaluation harness in view because I did not want a perfect-looking dashboard to carry an undefined claim. On its five-case fixed golden set, the offline harness reports precision 1.0 and recall 1.0: two true positives, three true negatives, no false positives, and no false negatives. That result belongs only to that fixed set. It does not establish a production accuracy rate, a universal safety rate, or a claim that every facade is safe.
The checks are specific. The documented-hazard analogues are flagged while the controls clear. A shared concave facade models at 102.6 degrees Celsius, compared with 74.6 degrees Celsius when made flat, and lower reflectance brings the flat reference to 64.0 degrees Celsius. The two flagged public-description approximations reach 122.3 degrees Celsius and 143.8 degrees Celsius, both above the 80-degree common-plastic heat-distortion consistency threshold. Those are reproducible demo checks, not measured outputs from the documented buildings.
I wrote those boundaries into the report because a result becomes less trustworthy when the scope is buried in a footnote. The value here is not that Caustic can pronounce a facade safe. Its value is that it makes a limited early decision inspectable, repeatable, and easy to challenge before a specialist has to unwind an unexamined concept.
The standard I now expect from an early engineering screen
I came away from this build with a narrower expectation for AI in architecture, and a more demanding one. I do not need an aesthetic system to reassure me that a glass form is compelling. I need a disciplined protocol that can bring deterministic physics into the concept loop, show the assumptions it used, and stop when the geometry outruns the model.
I built Caustic around a small but consequential distinction: a screen can guide where work goes next without impersonating the final work. Crescent Tower makes the focused-light risk visible during design iteration. The remediated variation shows that form and material choices can change the tested outcome. Helix Pavilion refuses the performance of certainty and sends the question to full optical simulation.
And if you would rather see it than read me describe it, here is the whole thing running end to end.
The runnable version will be published at https://veriprajna.com/demos/ai-architecture-structural-engineering so people can inspect that sequence rather than take the conclusion on faith. The design question I keep returning to is simpler than the interface: a rendering has not earned the authority of an engineering conclusion until its physical path of light has been tested.
Top comments (0)