Produsa build engine · internal demo

Same prompt, two engines

One identical brief built by Produsa's V5 (live today) and the new V7 pipeline. Open each and compare — speed, page weight, imagery accuracy, truthfulness, and design. Measured with real Lighthouse 12 (mobile).

Prompt: "Website for Anna's Kitchen, a South Indian restaurant in Indiranagar, Bangalore. We do dosas, idli, filter coffee and thalis. People should be able to see the menu and order on WhatsApp. Open 8am to 10pm daily."

V5 live today

  • Performance (mobile)40
  • Largest paint (LCP)10.1 s
  • Page weight1.43 MB
  • Images13 inline (base64)
  • Invented factsyes
  • Brief coverage60%
  • Build time~17 min
  • Design rubric7.7 / 10
Open V5 build →

V6 internal

Not shown here — V6's build driver didn't complete in this eval sandbox (headless-browser time/memory budget). It builds fine in production; it's just not reproducible in this harness. The meaningful comparison for "better/faster" is V5 (what's live) vs V7 (what's next).

build blocked in eval env

V7 new

  • Performance (mobile)98
  • Largest paint (LCP)1.8 s
  • Page weight17 KB
  • Images8 cached webp files
  • Invented facts0
  • Brief coverage100%
  • Build time~40 s
  • Design rubric8.1 / 10
Open V7 build →

What changed in V7: it researches the real menu and writes a precise image prompt per slot (authentic dosa / idli / filter coffee — not generic stock), then a vision judge drops any mismatch before ship. Images are cached files, not base64 inlined into the HTML, so the page is 84× lighter and paints in 1.8 s instead of 10 s. It binds only to the facts in the brief (0 invented), covers 100% of what was asked, and builds in seconds, not minutes — all while scoring higher on the design rubric.

Private demo · noindex · not a production URL · Produsa Engine V7 is behind a flag, not live. The Google-Maps embed on the Contact page needs a live domain to render.