𝐀 𝐬𝐞𝐦𝐚𝐧𝐭𝐢𝐜-𝐟𝐢𝐫𝐬𝐭 𝐅𝐢𝐠𝐦𝐚-𝐭𝐨-𝐜𝐨𝐝𝐞 - 𝐧𝐨𝐭 𝐣𝐮𝐬𝐭 𝐚 “𝐧𝐨-𝐜𝐨𝐝𝐞,” 𝐛𝐮𝐭 𝐢𝐧𝐭𝐞𝐧𝐭-𝐚𝐰𝐚𝐫𝐞 𝐜𝐨𝐝𝐞 𝐠𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐨𝐧. 𝐌𝐲 𝐢𝐝𝐞𝐚.
Hi Everyone!
Two years ago, working on a quant research prototype, I read the challenge of two separate worlds and visions - designers with their creation and developers focused on their logic - both ever conflicting on requirements.
The gap between intention and reality in our quest to build a prototype was enormous, cost months of iteration cycles and often just tedious work. That realization lead me to co-found CanvasEight.io and we've spent the last 18 months obsessing over this problem.
Initially we thought the challenge was just visual design fidelity in conversion to code.
We faced very challenging AI variability with just a 20%-40% fidelity accuracy in the first 6 months and reaching just only recently >70% Figma to code conversions.
We had done an extensive amount of work building a reliable AI structure, Machine Learning models which we've discarded or re-adapted over time and even an hybrid LLM environment scoped specifically for design to code generation.
Lately, we've arrived at something notably different and intriguing. We finally understood after countless design handoffs that the best teams weren’t just following specs, they were intuitively interpreting design intent.
That’s the essence of semantic reasoning and understanding not just replicating visuals, but grasping the functional logic beneath them.
From initially building towards high-fidelity design patterns, we transitioned trained our models for pattern recognition for UI components, automatic accessibility, state management inference.
We're looking at more complex logic, brand-specific patterns and performance optimization.
We believe AI should be a sophisticated starting point, handling >70% of implementation automatically while freeing teams to focus on what truly differentiates their product.
Often the real bottleneck is communication and we've seen this when designers test ideas, they refine their work faster. When developers focus on core logic instead of translating visuals, they build more robust systems.
Today AI-generated code has its limitations; edge cases still require human intuition and architecture decisions demand strategic oversight and good design judgment can't be automated away.
We’re now looking at our next phase: "AI-assisted craftsmanship" (I love coining our terms of reference); where automation handles routine implementation as humans creativity and oversight remain central.
The future is now beyond faster design-to-code translation and it’s transforming design and development into an ongoing conversation, closing the gap between what we design and what users actually experience.
Regards,
Marcello
Co-founder & CEO, CanvasEight.io
canvaseight figma
Comments
Post a Comment