CSS3

Responsive layouts for phones, tablets and desktops, using Tailwind and component themes.

In practice

An online grocery store with live payment updates

A grocery shop with a Next.js storefront and admin area and a NestJS API, using QR bank payments, live order and payment updates, and push notifications.

Problem: Customers pay by scanning a bank QR code, so the storefront has to learn when the payment arrives without the customer refreshing. Staff need to see new orders as they come in.
Contribution: Sole developer: storefront, admin area, API, database schema, payment and notification integrations, and the PM2 production setup.
Approach: A Next.js storefront and admin area built with Tailwind and Recharts. It talks to a NestJS API with modules for products, cart, orders, promotions, payments, notifications and site settings, using Prisma over PostgreSQL. Payments create VietQR/PayOS QR codes. When a payment is confirmed, a Socket.IO gateway pushes the status to that checkout session's room and alerts an admin feed room. Firebase Cloud Messaging sends push notifications. The API runs under PM2 in cluster mode.
Outcome: A working storefront and admin dashboard with QR checkout, live payment status, promotions and push notifications. The frontend is deployed on Vercel.

A multi-tenant point-of-sale platform

A point-of-sale product for small shops — a checkout app that keeps selling offline, an owner dashboard, a platform-admin console and subscription billing — built and deployed end to end.

Problem: Small shops need to ring up sales, track stock and close shifts on whatever device is at the counter, and they cannot stop selling when the connection drops. Each shop's data must stay isolated on one shared backend, and the platform has to bill shops for their plans.
Contribution: Sole developer. I designed and built the Angular checkout and admin apps, the NestJS API, the PostgreSQL schema, the realtime and offline sync, the billing and payment integrations, the CI pipeline and the EC2 deployment.
Approach: An Nx monorepo holds three apps: an Angular point-of-sale web app with a service worker and IndexedDB, an Angular platform-admin console, and a NestJS API on Fastify with Prisma over PostgreSQL. Tenancy is enforced in middleware that derives the shop from the access token, while the platform-admin plane uses a separate secret and never gets a shop context. Redis backs BullMQ jobs for work such as invoice renewal. Sales events reach open devices over WebSockets. Production runs on EC2 behind nginx with PM2, TLS from Let's Encrypt, and PostgreSQL on RDS.
Outcome: The product covers checkout, returns, inventory reservation, shelf-life lots, shifts, sales analytics, receipts, store settings, staff, VietQR and SePay payments, plan subscriptions with auto-renewal, and a platform-admin console. CI lints, tests and builds affected projects, enforces a bundle budget, scans for CVEs with Trivy and runs Playwright end-to-end tests.

Werewolf 3D: a server-authoritative party game

A ten-seat social deduction game played around a 3D campfire in the browser, with private lobbies, built-in AI players, text chat and a WebRTC voice path.

Problem: A Werewolf match needs ten players, and every player must see only what their role allows. Rejoining after a drop must not leak or lose state. A group of friends should be able to play without filling every seat, and without paid AI services.
Contribution: Sole developer: game rules, server, AI players, 3D client, networking protocol, media service and CI.
Approach: A TypeScript monorepo separates a pure rules package, a shared protocol package, a WebSocket game server, a Three.js client and a mediasoup media service. The server is the only authority. It projects a per-recipient view of each match, so hidden roles and night actions never reach clients that should not see them. AI players run inside the server and deliberately have no access to the match's private state — they act only on what they observed. Voice uses a mediasoup SFU, with TURN credentials minted per session from an HMAC of a shared secret.
Outcome: Playable locally and over a LAN with one to ten humans and up to nine AI players, core and starter role presets, reconnect handling with 60-second seat holds, and phone and tablet layouts. Voice is implemented but disabled by default; hardware benchmarks and human playtests are still pending.