Ecom AI

Blog content in BigCommerce stores is HTML with embedded links, images, tables, and custom formatting. Sending raw HTML to the model and asking it to improve the writing is a reliable way to get mangled markup back. Instead, a regex pass extracts only text nodes before the model ever sees the content. The model rewrites text, and the original HTML structure is reassembled around it. No markup is ever at risk.

Image and video generation have completely different timing constraints. Image renders complete in seconds: the route inserts a pending row, awaits a Supabase Edge Function call directly, and FAL writes the result back before the response returns. No queue needed. Video renders take 1–5 minutes and blow past the serverless timeout, so those jobs flow through Upstash QStash: QStash delivers the job durably to a worker route, which submits to fal's async queue, and an ED25519-verified webhook fires on completion. The same principle (don't queue what's fast, don't block on what's slow), but two different implementations.

Generic AI output looks generic because the model has no real brand context. The tool solves this at multiple layers: a global brand voice config, mood boards that define a visual aesthetic, and per-product notes for specific selling points or voice tweaks. Every generation call merges these three layers before building the prompt, so output is consistent with the actual brand and specific to the actual product, not a template filled in with product names.