01
Taking a Lovable-Built MVP to Production
The original prototype was created in Lovable and combined multiple independently implemented workflows and serverless functions. It proved the product idea, but tightly coupled logic and inconsistent processing flows made reliable operation and further development difficult.
We audited the existing implementation, completed missing functionality, aligned disconnected flows, and reinforced the critical path from product ingestion to publication. This transformed a functional proof of concept into a stable platform suitable for a real product catalogue and continuous product development.The existing product included numerous interconnected user journeys, from property discovery to listing management and communication.
02
Orchestrating a Multi-Stage AI Content Pipeline
The platform had to discover and classify source pages, extract structured product data, collect relevant supporting information, generate content, validate it, and prepare the result for publication.A real estate marketplace serves several audiences with different goals: buyers, tenants, property owners, and industry professionals.
These operations had to run reliably within the constraints of serverless functions and external AI services. We structured the workflow into clear stages with explicit statuses, fallbacks, and review points, making long-running processes easier to control and recover.
03
Making AI-Generated Content Safe to Publish
Product descriptions and FAQs must remain accurate, relevant, and consistent with the source data. A fluent response alone was not enough.Huawei devices require alternatives to several standard Google services. Push notifications, location features, and other platform-dependent functionality therefore could not rely on a single implementation.
We introduced layered quality control that combined prompt refinement, a separate LLM-based review pass, deterministic text checks, and human approval. Prompts were organized by purpose and versioned so the team could improve individual generation stages without disrupting the wider workflow.
04
Adding AI Video to the Product Content Flow
The platform needed to generate product videos without turning video production into a separate manual process.The platform continued to grow throughout the development process. New services, monetization models, and user scenarios had to be introduced alongside the ongoing modernization of existing functionality.
We connected script generation, avatar selection, voice configuration, background selection, rendering, preview, approval, and publication in one guided workflow. Retail teams could create a product-specific video using the same data already collected for SEO content.
05
Integrating With Existing Storefronts
Retailers needed to enhance existing product pages without migrating their e-commerce systems or rebuilding their catalogues.The platform needed to generate product videos without turning video production into a separate manual process.The platform continued to grow throughout the development process. New services, monetization models, and user scenarios had to be introduced alongside the ongoing modernization of existing functionality.
We implemented a storefront integration through Google Tag Manager. A dedicated tag allows approved descriptions, metadata, FAQs, and media to be injected into existing product pages while retailers retain control over what appears on their websites.