4. Comprehensive Reflection

• What worked well?
  The multi-tiered AI pipeline accelerated the initial conceptualization phase by connecting two separate areas: raw climate data and complex geometric forms. Using a Research AI to extract precise solar data for Kajang ensured the design was grounded in geographic reality, preventing the Text AI from generating completely impractical concepts. The Text AI worked excellently as a semantic bridge, taking engineering constraints (cavity depths, shading coefficients) and translating them into clear architectural logic ("living digital mashrabiya"). Finally, the Image AI was highly effective at instantly translating these abstract geometric definitions into a clear visual representation, which helped communicate the massing form and the light-transmitting quality of the honeycomb facade clearly.

• What failed?
  The primary failure was the lack of structural logic and physics awareness within the Image AI engine. Midjourney repeatedly failed to understand the load-bearing requirements of an architectural envelope. It generated configurations where structural steel grids seamlessly warped around razor-sharp corners without any corner mullions, expansion joints, or structural tie-backs to the internal floor plates. Furthermore, early image iterations suffered from "hallucinated organicism"—the honeycomb cells mutated into irregular, chaotic biological patterns resembling soap bubbles or spiderwebs rather than buildable, planar glass panels. The AI also routinely failed to render the physical depth of the double-skin cavity correctly unless forced through highly technical prompt overrides.

• Which AI tool was most useful?
  The Research AI (Perplexity) proved to be the most critical foundation for the workflow. While Image AI produces spectacular visual outputs, those visuals are functionally useless in an architectural presentation if they lack environmental and structural justification. By supplying the exact solar angles, thermal constraints, and cavity requirements specific to the Unipark site, the Research AI gave the entire project its logic and validity. It provided the objective engineering data necessary to transform a purely aesthetic choice (the honeycomb pattern) into a highly justifiable performance response to tropical solar heat gain.

• Where was human judgment required?
  Human judgment was required at every intersection between the AI tools to filter out non-viable solutions and correct errors:

Phase

AI Output Error / Limitation

Human Critical Intervention

Research Phase

Recommended mechanical, moving kinetic honeycomb panels based on temperate European standards.

Rejected moving parts due to high humidity and maintenance costs in Malaysia; converted design to a statically optimized, variable-depth honeycomb system.

Conceptual Phase

Generated unrealistic spatial layouts that compromised structural logic for poetic forms.

Restructured the program to align with structural grids, placing heavy double-volumes on targeted floor zones.

Visualization Phase

Rendered impossible structural cantilevers and floating glass skins without visible structural supports.

Reintroduced traditional load-bearing logic, mapping out internal concrete cores and columns behind the glass facade.

• 
  How did the workflow improve efficiency?
  This workflow compressed the standard conceptual and schematic design phases from weeks into a matter of days. Traditionally, conducting a deep micro-climate analysis, writing a comprehensive design brief, testing massing deformations, and creating high-quality 3D concept renders requires separate software pipelines (Rhino/Grasshopper, climate analysis plug-ins, rendering engines like V-Ray) and significant production time. The AI workflow allowed rapid design exploration, enabling the testing of multiple variations of massing forms and honeycomb scales within hours. This provided an immediate visual and narrative feedback loop that streamlined the design process.

• What are the risks of depending too much on AI?
  The primary risk is the erosion of structural, tectonic, and statutory discipline in architectural design. Because Image AIs generate highly convincing, polished, and beautifully lit renderings, designers face the danger of falling in love with a visual output that is structurally impossible, financially unfeasible, or completely non-compliant with building codes (such as Malaysia's UBBL fire safety and egress laws). Over-reliance on AI breeds a superficial approach to architecture, where form is completely disconnected from real-world construction limits, material thicknesses, structural load transfers, and spatial functionality.