Cabox — AI-Powered E-Commerce Platform
An online store where a single AI call identifies a product, estimates a market-competitive price, and writes a bilingual listing — from one product photo
Running an online store means listing hundreds of products — each needing photos, descriptions, prices, specs, and marketing copy. Most store owners do it by hand or hire freelancers. The bottleneck isn't inventory — it's the work of turning a product into a listing.
I built Cabox so one Gemini 2.5 Flash call collapses most of that workflow. Take a photo. The model identifies the product, estimates a market-competitive price and converts it to Costa Rican colones, writes Spanish and English names and descriptions, generates technical specs, an auto-generated SKU, a URL slug, and a category assignment — and a Google Custom Search pass (with a scraper fallback) pulls promotional images matching the product. All from that one image, in a single pass.
The Problem
The hardest part of e-commerce isn't building the store — it's populating it. Every product needs:
- A name and description (in Spanish and English for a Costa Rican audience)
- An accurate market-competitive price
- Technical specifications
- Promotional copy for Facebook, Instagram, and WhatsApp
- Product images for listings and ads
Doing this manually for hundreds of products takes weeks. Hiring freelancers is expensive and inconsistent. The result is that stores launch with thin listings, weak pricing, and no promotional strategy — which means fewer sales.
The Approach
I built a Next.js 16 storefront with a full admin panel, then layered a single AI model — Gemini 2.5 Flash — on top. The core flow runs in one pass:
- The admin uploads a product photo (or multiple photos from different angles).
- Gemini 2.5 Flash analyzes the images — it identifies the product visually, estimates a market-competitive price, converts it to Costa Rican colones, and returns a complete product package: bilingual names, marketing descriptions, technical specs, an auto-generated SKU, a URL slug, and one of eight predefined categories. The model also returns a confidence level (high/medium/low) so the admin knows where to focus review.
- The AI suggests a selling price at exactly 75% of the estimated market price — giving the owner margin while staying competitive. This ratio is enforced in the prompt and hard-computed in code, so it never drifts.
- A Google Custom Search pass (with a googlethis fallback) then finds promotional images matching the product — run on the search query Gemini returned, so the images are relevant to what the model just identified.
- The admin reviews and can edit any field — the AI fills, the human decides.
Once the product is live, ad copy is generated on demand. The admin clicks "WhatsApp", "Facebook", or "Instagram" and a second Gemini 2.5 Flash call writes copy tailored to that channel's tone and format — always in Costa Rican Spanish, with the product URL included. The ads are copied to a clipboard for manual posting; the platform doesn't auto-publish to social platforms.
The storefront itself handles the rest: five payment methods (Stripe, PayPal, SINPE Móvil, bank transfer, cash), bilingual Spanish/English via next-intl, and an installable PWA manifest so the store can be added to a phone home screen.
The Outcome
What used to take 30–60 minutes per product — photograph, research, write descriptions in two languages, find images — now happens in a single AI-assisted pass. The owner uploads a photo, reviews the AI's output, edits where needed, approves, and the listing is live.
The store is live at cabox.dolcebuy.com with real products, real payments, and real customers. The pipeline doesn't just accelerate listing creation — it enforces a consistent pricing discipline (the 75% rule) and ensures every product ships with professional bilingual content that would cost thousands to produce per batch through freelancers.
One caveat worth naming: pricing is the model's estimate from its training knowledge, not a live internet price scrape. For commodity products with well-known Amazon prices this is strong; for niche or volatile-priced inventory the admin still needs to sanity-check. The platform is honest about this — the AI suggests, the human decides.
Key Takeaway
The AI pipeline isn't the feature — speed-to-market is. A store that can list products, price them consistently, and promote them in minutes instead of hours doesn't just save time. It sells more. One model, four AI-powered features, one honest workflow: the AI drafts, the human decides.
Ready to Build Your Platform?
One 30-minute call to see if we're a fit. No pitch. No pressure. Just a conversation about what you need to build.
Most engagements start with a bounded architecture diagnostic: $5,000 to $12,000, 1 to 3 weeks. Fixed-price platform builds: $20,000 to $80,000, milestone-gated, 50% deposit. Fractional CTO: $6,000 to $10,000 per month, 90-day minimum.