Jelma: An Indonesia-First AI Website Builder

The problem

Fewer than 2% of Indonesian online sellers have their own website, while around 95% sell through WhatsApp and Instagram. Global AI builders are English-only, dollar-priced, and card-only; local alternatives are template pickers or developer-facing tools. Jelma, our own product under active development, puts the non-technical UMKM owner first and the engineer second.

What we are building

• Story to website. A guided intake in Bahasa, formal or gaul, turns a business description into a complete static site: an Indonesian block library, a WhatsApp CTA by default, mobile-first with a 360px Android floor.
• Design system first. Every project starts from a curated preset; applying or switching a design system costs no credits, so consistency is free.
• Edit by click or by chat. Common small edits (text, colour, spacing) apply in seconds with no LLM call and no credits.
• Publish like a local. A Jelma subdomain or your own domain on Cloudflare static hosting, prepaid in Rupiah via QRIS, virtual account, or e-wallet (iPaymu), with a transparent credit ledger.
• Abuse-resistant by design. Nothing gets a public URL without passing both gates (a generation-time LLM guard and a publish-time scan), OTP verification before the first publish, and a per-site kill switch.

Engineering decisions

Spec-driven end to end. A PRD with 25 features and 15 locked decisions is decomposed into 10 epics and 41 stories; every story is built test-first, and proven conventions are promoted into guard tests that must be mutation-verified before they are trusted. The database is locked to MariaDB as decision D1.

AI cost made explicit. The credit ledger and model-cost multipliers are published to users; edits that do not need a model genuinely do not call one.

Indonesian compliance from day one. PSE, UU PDP, and takedown SLAs are part of the design, not patched in after launch.

Status

Jelma has not launched. The application foundation, auth, payments, and its spec-driven build pipeline run; the remaining epics are being worked story by story. We list it here as it is: a study in how we build an AI product for the Indonesian market, not a finished product.