2025 – Present · Fullstack Architect & Creator · Full-stack · Building the product
Puchi — Vietnamese Language Learning Platform
Next.js PWA with 5 Go microservices (Kratos v3), Limen sessions, NATS events, and ArgoCD GitOps
What this system is for
Puchi is a learner-facing language platform. Its architecture starts with the learner journey: stories, scenes, activities, progress, pronunciation media, reminders, and the ability to evolve the curriculum without rewriting learner history.
What I owned
Product architecture, Next.js PWA, Go services, learning domain model, authentication, event flows, media delivery, and GitOps deployment.
Overview & Context
Puchi is a learner-facing language platform. Its architecture starts with the learner journey: stories, scenes, activities, progress, pronunciation media, reminders, and the ability to evolve the curriculum without rewriting learner history.
Problem & Constraints
Online language learners need fast, offline-capable PWA experiences with instant lesson progress synchronization, audio pronunciation streaming, and resilient microservice APIs under traffic spikes.
System Architecture
User ──► Cloudflare CDN ──► Next.js PWA (Limen sessions) ──► Envoy Gateway
│
▼
Go Services (Kratos v3) ──► PostgreSQL / NATS / Cloudflare R2
[auth, core, learn, media, notification]
│
▼
K3s Cluster (v1.35) + ArgoCD GitOps + Traefik Ingress- Next.js PWA with Limen authentication session management
- 5 Go Microservices (Kratos v3) with NATS asynchronous event messaging
- Cloudflare R2 object storage for audio pronunciation assets
- ArgoCD GitOps deployment pipeline on multi-node K3s cluster
Engineering Deliverables
- Built responsive Next.js PWA frontend with progress tracking, interactive quiz exercises, and audio pronunciation streaming via Cloudflare R2.
- Developed 5 Go microservices monorepo (app/auth, core, learn, media, notification) powered by the Kratos v3 framework.
- Implemented asynchronous domain events via NATS (auth.user.created, learn.lesson.completed, email.send).
- Automated continuous delivery via ArgoCD GitOps on a 3-node K3s cluster with Traefik ingress and Cloudflare Tunnels.
Engineering Signals & Outcomes
- 5 Go Microservices (Kratos v3) + NATS Event Bus
- Next.js PWA with Limen sessions & Cloudflare R2
- Declarative ArgoCD GitOps on K3s Kubernetes
Engineering Outcomes
- A product model that starts from learner actions rather than service count.
- Separated curriculum content, learner state, media delivery, and notification work.
- A repeatable path from product change to a declared cluster deployment.
Engineering threads behind the project
Each thread links to notes focused on one concrete decision or engineering problem.
Modeling learning before modeling services
The Story → Scene → Content → Activity model separates curriculum structure from attempts, progress, XP, and streak state.
Authentication that can be revoked centrally
Why Puchi uses opaque sessions and internal introspection instead of spreading raw JWT handling across the product.
A delivery path designed for a small product team
GitHub Actions, GHCR, ArgoCD, and K3s turn a product change into a repeatable deployment without manual server access.
Technologies & Tooling
Go · Kratos v3 · Next.js · PWA · NATS · PostgreSQL · ArgoCD · K3s · Cloudflare R2