FRM← back to projects04 / 05

04 / CASE STUDY

PRODUCT EXPERIMENT

AI Sales Craft.

From a raw briefto JSON ready to render.

A dual-LLM sales page generator whose output is JSON, already past schema validation and ready to render.

STATUS
EXPERIMENT
INPUT
BRIEF
OUTPUT
JSON

BUILT WITH

  • Next.js
  • Express
  • TypeScript
  • PostgreSQL
  • Vercel

01 / PROBLEM

Free-form outputcannot be rendered.

Raw product details (name, features, audience, pricing, the selling point) have to become a structured sales page, and doing that by hand is slow. Loose model prose does not solve it. What was needed was output with a fixed shape, consistent enough to render across several visual templates.

02 / APPROACH

Schema first,model second.

The backend prompts the LLM to return strict JSON matching a defined schema, then validates it before storing. Groq with Llama 3.3 70B is the primary engine for speed; when Groq is unavailable or errors, the request automatically falls back to Gemini under the same schema contract. The frontend maps one generation onto three distinct templates.

  1. 01BRIEF
  2. 02GROQ
  3. 03SCHEMA CHECK
  4. 04GEMINI (FALLBACK)
  5. 05THREE TEMPLATES

VALIDATED

03 / LEARNED

The fallback camefrom real pain.

This project started on Laravel 11 + MySQL. When deployment moved to Vercel, the backend was rewritten in Express + TypeScript with PostgreSQL on Neon. That full rewrite taught me more about both ecosystems than either version alone. The fallback arrived later, out of incidents: free-tier LLM APIs fail often, and handling that gracefully matters more than the happy path.

SCHEMAJSON that passes schema validation.
FALLBACKGroq → Gemini, one output contract.
HISTORYJWT-authenticated CRUD with search and pagination.

NEXT PROJECT

Vidio Belajar05 ↗