AI Trade-Show Kiosk + Shopify Engagement
AI Image Pipeline, Headless Checkout, and Full Shopify/Klaviyo/Recharge Delivery
Two-device trade-show kiosk built for an off-roading e-commerce client. The AI image pipeline turns a customer photo and a scene prompt into a print-ready 51.4" light-bar insert and a Shopify draft order. Same engagement covered the client's full Shopify operation — admin audit, lifecycle email, subscription recovery, structured-data SEO, and Meta catalog repair.
- Four-model AI image pipeline (gpt-image-1.5, gpt-4o-mini, Claid 4× upscale, Replicate rembg) — print-ready 15,420 × 1,320 px output
- 7-level graceful-degradation fallback chain so the kiosk never returns “generation failed”
- 8 switchable prompt strategies + opentype.js text burn-in (zero AI-misspelling risk on print files)
- Two-device session sync (customer phone + attendant tablet) via QR codes and Postgres polling
- Headless Shopify draft-order creation + 61-section Shopify admin audit and remediation
- 3 Klaviyo lifecycle flows + Recharge subscription recovery + Meta catalog repair across 344 theme files
- Next.js 16
- React 19
- TypeScript 5
- Sharp
- Claid.ai
- OpenAI gpt-image-1.5
- Klaviyo
- Recharge
- Vercel Pro
- Shopify Admin GraphQL
- Neon Postgres
- Drizzle ORM
Two deep-dives into the AI surface
The kiosk is an AI feature wrapped in a UI — these are the parts worth showing. Each case study is a focused walkthrough — architecture, the engineering 'why,' and real annotated source. Each is independently linkable, so you can share the one that matches the role.
AI Image Pipeline
Customer photo + scene description → print-ready 15,420 × 1,320 px panoramic in ~45s, with a 7-level graceful-degradation chain so the booth never tells a paying customer "generation failed."
Read case study →Prompt Engineering
Image models trained on 1:1 / 16:9 data don't have a useful prior for 11.68:1. The prompt has to do composition work the model wasn't trained for — so I built a switchable harness with eight strategies and shipped the one with the most stacked constraints.
Read case study →Want to see the harness or the diagnostics object?
Happy to walk through the 8-prompt comparison run, the per-phase ms timings, or the full 7-level routing decision on a call.