insonl
← all work

Manh Tri

Manh Tri sells wood and vinyl flooring, bamboo-charcoal wall panels and outdoor decking in Vietnam. We rebuilt manhtri.com as a fast Next.js frontend on top of the company's existing WordPress catalogue, and added a Vietnamese-language chat assistant that searches products, articles and finished installations before it answers.

Year
2026
Category
E-commerce & AI
Stack
TypeScript, PHP
Manh Tri

About the project

The Manh Tri team already kept its products, articles and photos of finished jobs in WordPress. Rather than replace that workflow, we put a Next.js frontend in front of it. Pages are server-rendered and cached, images are served as WebP, and the cache is purged when content changes, so staff keep editing where they always have.

Where AI is used

The site has a chat assistant, “CSKH Manh Tri 24/7”, built on Claude through the Anthropic SDK. It does not answer from memory. It calls tools that search the live product catalogue, the how-to articles and the gallery of completed installations, then shows product and project cards under its reply.

  • Grounded answers: the instructions forbid inventing prices, product codes or policies; prices are shown as reference values and the customer is invited to call or message on Zalo for a quote.
  • Staff handoff: when a customer needs a survey, estimate or order, the assistant points them to the hotline and Zalo.
  • Cost control: the assistant runs on a single small model, with per-session, per-visitor and daily token budgets.

Key features

  1. Instant search: a search index with live suggestions and clear result counts.
  2. Shopping flow: category archives with sorting and paging, a calculator, accessories and past projects shown alongside products.
  3. Mobile and accessibility: sticky search, focus-trapped menus that close with Back, and a trust strip with popular categories.
  4. SEO for people and for AI search: structured content, redirects from the old .html URLs, and llms.txt files describing the site to AI crawlers.
  5. Visitor analytics and a deploy setup with Docker and a cache warm-up script.

A second build

A separate monorepo explored the same brand on Cloudflare: a bilingual Vietnamese and English site with an admin panel, hybrid search over vector embeddings, and an AI chat assistant.

Technical aspects

  • Frontend: Next.js with server rendering and incremental static regeneration
  • Content: WordPress catalogue read over its API, with a cached search index
  • AI: Claude via the Anthropic SDK with tool calling
  • Delivery: Docker, Traefik and Cloudflare cache purge

next case

ChatSQ8

ChatSQ8