How This Site Is Built — Behind the Scenes of KoishiAI

TL;DR: KoishiAI achieves near-zero monthly costs of approximately $1.50 by running a full AI content pipeline on a home PC with an RTX 5090. The architecture combines Astro 5 for static frontend generation, Firebase Spark for free hosting, and local Ollama instances running Qwen3.5 and Qwen3.6 models for autonomous article creation.

Key facts

• The site runs on Astro 5 with Tailwind CSS and TypeScript strict mode, generating pure static HTML with no server runtime.
• Hosting utilizes the free Firebase Spark plan, providing 10GB storage and CDN edge nodes in Bangkok.
• AI processing relies on local Ollama instances running qwen3.5:27b for general tasks and qwen3.6:35b-a3b for long-form content.
• Hardware requirements include an AMD Ryzen 9 CPU, 64GB RAM, an RTX 5090 (32GB VRAM), and an RTX 5080 (24GB VRAM).
• Total monthly operational costs are approximately $1.50 USD, covering only electricity for generating roughly two articles per day.
• The automated pipeline uses FastAPI, SQLite, and APScheduler to orchestrate agents for scouting, writing, image sourcing, SEO optimization, and translation.
• Astro was selected over Next.js to reduce the attack surface, avoid server-side vulnerabilities, and eliminate React hydration overhead for lower LCP.

This post walks through how KoishiAI is built — from stack choices to making AI autonomously write and publish articles — at an effective cost of near zero per month.

The stack

Frontend

• Astro 5 for static site generation — output is pure HTML, no server runtime
• Tailwind CSS — utility-first
• TypeScript strict mode for safety
• IBM Plex Sans Thai + Inter — free fonts that render beautifully in both languages

Hosting

• Firebase Hosting Spark plan (free) — 10GB storage, CDN edge in Bangkok
• Not using Firebase Storage (now requires Blaze) — images live in Hosting directly

Metadata DB

• Firestore Spark plan (free) — article metadata and dashboard queries

AI (all runs on a home PC)

• Ollama + qwen3.5:27b for general work
• qwen3.6:35b-a3b for long-form pieces
• ComfyUI for custom-styled images
• Pexels/Unsplash APIs for stock illustrations

Orchestration

• FastAPI (Python) — dashboard backend
• SQLite for the queue + logs
• APScheduler for cron jobs
• Firebase Admin SDK to push data to the cloud

The article pipeline

New topic from RSS feed
    ↓
topic_scout agent filters and ranks candidates
    ↓
content_writer drafts markdown
    ↓
image_finder sources illustrations (Pexels / Unsplash / ComfyUI)
    ↓
seo_optimizer tunes meta tags and keywords
    ↓
translator produces the other language
    ↓
editor proofreads
    ↓
publisher: write .md → git commit → astro build → firebase deploy
    ↓
fb_poster shares to Facebook Page ~10 min after deploy

All of this runs on one home PC — no foreign APIs required.

Hardware

• CPU: AMD Ryzen 9
• RAM: 64GB
• Primary GPU: RTX 5090 (32GB VRAM) for large LLMs
• Secondary GPU: RTX 5080 (24GB VRAM) for parallel image generation
• Storage: 2TB NVMe

Electricity for producing 2 articles per day averages around 50 THB ($1.50 USD) per month.

Real monthly costs

Why Astro instead of Next.js

1. Much smaller attack surface — Astro has no middleware, no server actions; the entire site is static HTML
2. Fewer critical CVEs lately — Next.js recently had a critical middleware auth vulnerability
3. i18n built-in — no third-party library needed
4. Lower LCP — no React hydration overhead

Source code

This project is part of the Koishi monorepo, which isn’t yet public — but we plan to gradually open-source it once the system stabilizes (estimated 1–2 months).

Questions about the stack or want to see a specific implementation detail? Drop a comment on our Facebook Page.