AI Blog Tech Stack

This stack writes my blog on my own hardware. Astro handles the static site, Mistral Small 4 writes the drafts, and ComfyUI FLUX generates the hero images, all running on a single DGX Spark. I curate and quality-check, the pipeline does the rest.

Quick Take

• One machine runs everything: Astro, Mistral, ComfyUI FLUX (DGX Spark with 128GB unified memory)
• Hero images are WebP at quality 82 (~20–160 KB) with LCP optimization
• Cloudflare Tunnel exposes the site without exposing ports (v2024.5.1)
• Lightning tips support the project directly via Alby integration

Stack in One Docker Compose

version: '3.8'
services:
  astro:
    image: withastro/astro:5.6.1-tailwind-ts
    volumes:
      - ./dist:/usr/src/app/dist
    ports:
      - "127.0.0.1:4321:4321"
    networks:
      - webshop_default
  nginx:
    image: nginx:1.25.5-alpine
    volumes:
      - ./dist:/usr/share/nginx/html
      - ./config/nginx.conf:/etc/nginx/nginx.conf
    ports:
      - "80:80"
    networks:
      - webshop_default
  cloudflared:
    image: cloudflare/cloudflared:2024.5.1
    command: tunnel --no-autoupdate run --token ${CF_TUNNEL_TOKEN}
    networks:
      - webshop_default
networks:
  webshop_default:
    external: true

Why this matters: no Node in production, nginx serves static files with gzip/brotli compression, Cloudflare Tunnel hides the origin. The DGX Spark handles Mistral and ComfyUI sequentially (CUDA 12.4, NVIDIA driver 550.54.15). 128GB unified memory is enough for FLUX.1-schnell but can hit OOM errors with larger batch sizes.

Gotcha: ComfyUI and Mistral can’t run at the same time on this hardware. Stop Mistral before generating images (error: CUDA out of memory when both services active). Use sudo systemctl stop mistral before image generation.

Watch Out: Cloudflare Tunnel v2024.5.1 may fail to reconnect after network interruptions. Monitor with journalctl -u cloudflared -f and set up systemd restart policy.

Content Pipeline from Markdown to Live Site

cd /data/projects/sovereign-blog
python3 scripts/update_blog_from_gitea.py --run-now --model mistral-small-4
python3 scripts/generate_blog_images.py --model flux-schnell --batch-size 4
npm run build -- --mode production

The pipeline reads markdown from src/content/blog/, feeds it to Mistral Small 4 via port 30000 (HTTP API endpoint), then generates WebP images with ComfyUI FLUX.1-schnell. Output goes straight into dist/, no rebuilds needed for content changes.

Gotcha: Protected articles (frontmatter protected: true) are skipped automatically. Delete them from processed_files.json to regenerate. The script logs skipped files to /var/log/sovereign-blog/skipped.log.

Watch Out: If Mistral API (localhost:30000) is unreachable, the pipeline fails silently with exit code 1. Verify service status with systemctl status mistral before running.

First-Hand Note: The --run-now flag triggers immediate execution. Without it, the script waits for the next scheduled run (cron at 02:00 UTC). I once forgot this flag and waited 24 hours for updates.

Frontmatter Schema That Powers the Pipeline

title: "Build a Self-Hosted AI Blog with Astro, Mistral, and ComfyUI"
description: "A complete pipeline for generating, optimizing, and deploying AI-written content on your own hardware"
date: 2026-04-22
tags: ["sovereign-ai", "self-hosted", "astrod"]
heroImage: /images/blog/build-self-hosted-ai-blog/hero.webp
img_score: 4.2
affiliate_links:
  - label: "Alby"
    url: "https://getalby.com"
    note: "Lightning tips for direct support"
protected: false

The schema enforces EEAT scores (1-5 scale), image quality, and affiliate links. img_score comes from rate_images.py, run it after every pipeline to keep scores current.

Gotcha: rate_images.py opens each hero image in a viewer. Press 1–5 to score, s to skip, q to quit. Scores write back to frontmatter automatically. If the viewer crashes (common with large images), check /tmp/image_rating.log for partial updates.

Watch Out: Frontmatter validation fails silently if YAML syntax is incorrect. Use yamllint in CI to catch errors early. I once had a colon in a tag value (tags: ["ai:ml"]) that broke the pipeline for hours.

Image Pipeline from FLUX to WebP in 82 KB

python3 scripts/generate_blog_images.py --dry-run --output-dir /tmp/test-images

ComfyUI FLUX.1-schnell outputs WebP directly at quality 82 using libwebp 1.3.2. No separate conversion step, pipeline handles it. Hero images land in public/images/blog/<slug>/hero.webp at ~20–160 KB, keeping LCP green.

Gotcha: If an image exceeds 400 KB, the Insights page flags it in red. Run convert_images_to_webp.py once to batch-convert legacy PNGs. The script uses cwebp -q 82 -mt for multi-threaded conversion.

Watch Out: FLUX.1-schnell can generate corrupted WebP files under memory pressure. Symptoms include Error: Failed to decode image in browser. Mitigation: Reduce batch size to 2 and monitor memory with nvidia-smi.

First-Hand Note: I once generated a 5MB WebP file due to a ComfyUI node misconfiguration. Debugging involved checking /var/log/comfyui/flux.log and comparing file sizes with ls -lh.

Insights Page Shows Pipeline Metrics

---
import { getBlogStats } from '../utils/blog-stats.ts';
const stats = await getBlogStats();
---

The /insights/ page shows EEAT scores, img_score, actual file sizes (heroKb), and whether a caricature exists. All metrics are build-time, no backend needed.

Gotcha: The KPI strip shows five hardcoded metrics. If you add more, update the strip in src/pages/insights.astro. The current metrics are:

1. Average EEAT score
2. Average image score
3. Largest hero image (KB)
4. Total articles processed
5. Pipeline success rate

Watch Out: The insights page may show stale data if the build fails. Check /dist/insights.json for the latest metrics. I once had a failed build that showed 0 articles for a week.

First-Hand Note: The metrics are generated during npm run build. If you modify the pipeline, verify the output with node scripts/verify-insights.mjs before deploying.

What I Actually Use

• Mistral Small 4 (v1.0.0): The only model I run locally, fast enough for drafts (avg. 12s per article). Uses 8GB VRAM on DGX Spark.
• ComfyUI FLUX.1-schnell (v1.1.4): Generates hero images without cloud APIs. Requires 16GB VRAM for batch size 4.
• Cloudflare Tunnel (v2024.5.1): Exposes the site without opening ports. Configuration stored in /etc/cloudflared/config.yml.
• Nginx (1.25.5): Serves static files with brotli compression. Config in /etc/nginx/sites-available/blog.conf.
• Astro (5.6.1): Static site generator with Tailwind CSS. Builds to /dist in ~30s.

Watch Out: Astro 5.x requires Node.js 18+. The DGX Spark runs Ubuntu 22.04 with Node.js 20.12.2 installed via nvm. Downgrading Node.js caused build failures until I pinned the version in package.json.

First-Hand Note: I initially tried running Mistral and ComfyUI in parallel but hit CUDA_ERROR_ILLEGAL_ADDRESS errors. The DGX Spark’s unified memory architecture isn’t ideal for multi-process GPU workloads. Now I use a simple bash script to alternate services:

#!/bin/bash
# /usr/local/bin/toggle-services.sh
case "$1" in
  mistral)
    systemctl stop comfyui
    systemctl start mistral
    ;;
  comfyui)
    systemctl stop mistral
    systemctl start comfyui
    ;;
  *)
    echo "Usage: $0 {mistral|comfyui}"
    exit 1
esac