blackroad-marketing/ads/social/dev-to-articles.md

BlackRoad Dev.to / Hashnode Cross-Posts

Principle: Authority + Reciprocity (give technical depth, earn credibility) Rule: These are developer audiences. Lead with code and architecture, not marketing.

---

Article 1: Setup Guide

Title: How I Run 16 AI Models on a Raspberry Pi 5 + Hailo-8 (Complete Setup Guide)

Tags: #ai #raspberrypi #selfhosted #tutorial

Body:

I run 16 language models on a Raspberry Pi 5 with a Hailo-8 AI accelerator. This isn't a benchmark post — this is the production setup I use daily for BlackRoad OS.

Total hardware cost: $180. Total cloud bill: $0/month.

Here's exactly how to set it up.

## Hardware

| Component | Price | Purpose |
|-----------|-------|---------|
| Raspberry Pi 5 (8GB) | $80 | Compute |
| Hailo-8 M.2 AI Kit | $99 | 26 TOPS neural inference |
| 128GB microSD (A2) | $15 | Storage |
| USB-C power supply (27W) | $12 | Power |
| **Total** | **$206** | |

Optional but recommended:
- NVMe SSD via Pi 5 HAT ($30-50) — faster model loading
- Ethernet cable — more reliable than WiFi for serving

## OS Setup

Flash Raspberry Pi OS (64-bit, Bookworm) with the Raspberry Pi Imager.

Enable SSH during flashing (set hostname, username, password).

Boot and SSH in:

```bash
ssh pi@<your-pi-ip>

Update everything:

sudo apt update && sudo apt upgrade -y

Install Ollama

One command:

curl -fsSL https://ollama.com/install.sh | sh

Verify:

ollama --version

Load Models

Pull the models you want. Here's what I run:

ollama pull llama3
ollama pull mistral
ollama pull codellama
ollama pull phi3
ollama pull gemma:2b
ollama pull nomic-embed-text

List loaded models:

ollama list

You can load as many as your storage allows. Ollama swaps models in and out of memory efficiently — only one runs at a time in RAM, but switching is fast.

Install Hailo-8 Drivers

Add the Hailo APT repository:

sudo apt install -y hailo-all

Verify the accelerator is detected:

hailortcli fw-control identify

You should see: Board Name: Hailo-8 and 26 TOPS.

Test Inference

Run a quick test:

ollama run llama3 "What is edge AI inference?"

For the Hailo-8 specifically, you can run object detection and classification models natively:

hailortcli run <model.hef>

Serving Over the Network

Ollama serves on port 11434 by default. To expose it to other nodes on your network:

OLLAMA_HOST=0.0.0.0:11434 ollama serve

Now any machine on your LAN can query:

curl http://<pi-ip>:11434/api/generate -d '{
  "model": "llama3",
  "prompt": "Explain edge computing in one sentence."
}'

What I Actually Run

In production, I have 5 Pis orchestrated with Docker Swarm:

• Alice: gateway, DNS, databases
• Cecilia: AI inference (this setup + Hailo-8)
• Octavia: git server + second Hailo-8
• Aria: agent runtime + NATS messaging
• Lucidia: web apps + CI/CD

Two Hailo-8s = 52 TOPS combined. WireGuard mesh encrypts everything. Cloudflare Tunnels provide ingress without opening ports.

This serves 30 websites, 50 AI skills, and 207 git repositories. Monthly cost: electricity (~$5-8).

The Economics

	Cloud GPU (H100)	Raspberry Pi + Hailo-8
Cost	$3.90/hr	$206 one-time
Annual cost	$33,696	~$70 electricity
Inference	65 TOPS	26 TOPS
Privacy	Vendor-dependent	Absolute
Vendor lock-in	Yes	No

The cloud GPU is faster. But for inference workloads that run 24/7, the Pi pays for itself in 53 hours of avoided cloud compute.

Questions?

I'm happy to answer anything about this setup. The full architecture is at blackroad.io.

The project is open and the infrastructure is real — this post was served through it.

---

## Article 2: Architecture Deep-Dive

**Title:** I Replaced My Cloud Infrastructure with 5 Raspberry Pis — Here's the Full Architecture

**Tags:** #architecture #selfhosted #devops #ai

**Body:**
```markdown
This is the complete production architecture for BlackRoad OS. Not a demo. Not a weekend project. The actual infrastructure serving 30 websites, 207 git repos, and 50 AI skills.

## Network Topology

Internet │ ├── Cloudflare (DNS, CDN, Tunnels, Pages) │ ├── 95+ Pages sites │ ├── 40 KV namespaces │ ├── 8 D1 databases │ └── 10 R2 buckets │ ├── Cloudflare Tunnel → Alice (.49) │ ├── Pi-hole (DNS filtering, 120+ blocked domains) │ ├── PostgreSQL │ ├── Qdrant (vector search) │ └── Gateway / reverse proxy │ ├── Cloudflare Tunnel → Cecilia (.96) │ ├── Ollama (16 models) │ ├── Hailo-8 (26 TOPS) │ └── Embedding engine (nomic-embed-text) │ ├── Cloudflare Tunnel → Octavia (.101) │ ├── Gitea (207 repos, primary git host) │ ├── Docker Swarm manager │ └── Hailo-8 (26 TOPS) │ ├── Cloudflare Tunnel → Aria (.98) │ ├── Agent runtime │ └── NATS v2.12.3 (pub/sub mesh) │ └── Cloudflare Tunnel → Lucidia (.38) ├── 334 web applications └── GitHub Actions runner

## Internal Mesh

All nodes connect over WireGuard:

Alice ←→ Cecilia ←→ Octavia ←→ Aria ←→ Lucidia ↕ ↕ ↕ ↕ WireGuard encrypted mesh (10.8.0.x)

NATS v2.12.3 provides pub/sub messaging between 4 nodes. Agents on any node can publish tasks and subscribe to results.

## Key Design Decisions

**Why Gitea instead of GitHub?**
Gitea is primary. GitHub is a mirror. We own the git server. If GitHub changes pricing, policies, or terms — our repos don't move. Gitea runs on Octavia and syncs downstream to 17 GitHub orgs every 30 minutes.

**Why Cloudflare Tunnels?**
Zero open ports on any Pi. All ingress goes through authenticated Cloudflare Tunnels. No port scanning surface. No firewall rules to maintain (except Lucidia, which runs UFW).

**Why Pi-hole?**
Network-level DNS filtering. 120+ tracking/advertising domains blocked for every device on the network. This isn't ad-blocking — it's a security boundary.

**Why Hailo-8 instead of cloud GPUs?**
52 TOPS for $198 total vs. $33,696/year for one cloud H100. The Hailo-8 handles classification, detection, and inference acceleration. Ollama handles language model serving. Together, they cover the full inference stack.

**Why NATS?**
Lightweight pub/sub messaging designed for distributed systems. Agents on different Pis communicate by publishing to topics and subscribing to results. No message broker to maintain. No Kafka cluster to babysit.

## Services Map

| Service | Node | Port | Purpose |
|---------|------|------|---------|
| Pi-hole | Alice | 53, 80 | DNS filtering |
| PostgreSQL | Alice | 5432 | Primary database |
| Qdrant | Alice | 6333 | Vector search / RAG |
| Ollama | Cecilia | 11434 | LLM inference (16 models) |
| Gitea | Octavia | 3100 | Git hosting (207 repos) |
| NATS | Aria | 4222 | Agent messaging |
| Web apps | Lucidia | various | 334 applications |

## Monitoring & Automation

Cron jobs on the Mac (Alexandria, .28):
- Every 5 min: health check + fleet collector
- Every 15 min: Cecilia model sync
- Every 30 min: website checks + git sync
- Daily 3am: backups + sovereign mesh check
- Every 6h: Google Drive sync (rclone)
- Every 12h: rsync to DigitalOcean droplet

## Replication

This entire architecture is reproducible. Total hardware cost: ~$400. Setup time: a few hours with the guide.

Full setup guide: [link to Article 1]

Questions welcome — happy to share configs, scripts, or troubleshooting for any of these services.

---

Article 3: The Psychology Post (Cross-post from blog)

Title: 80% of Advertisements Are Misunderstood — I Studied Why, Then Built a Marketing System That Can't Be

Tags: #marketing #psychology #transparency #startup

(Cross-post blog/02-advertisements-misunderstood.md with Dev.to formatting)

---

Publishing Schedule

Week	Article	Platform
1	Setup Guide	Dev.to + Hashnode
2	Architecture Deep-Dive	Dev.to + Hashnode
3	Psychology of Advertising	Dev.to
4+	Cross-post new blog posts as they publish	Both

Dev.to-Specific Rules

• Use their canonical URL feature to point back to blackroad.io/blog (SEO benefit)
• Include a "series" tag to link articles together
• Respond to every comment within 24 hours
• Don't use cover images that look like marketing — use terminal screenshots or architecture diagrams