feat: LinkedIn, Dev.to, onboarding emails, landing page

LinkedIn: 7 posts (founder story, psychology, market thesis, culture, contrarian, sovereignty, education) Dev.to: 3 articles (Pi setup guide, architecture deep-dive, psychology cross-post) Onboarding: 3-email post-deploy sequence (welcome → expand → identity) Landing page: 'Own Your Stack' full page copy with A/B test plan RoadChain-SHA2048: acf9a01550f3cb91 RoadChain-Identity: alexa@sovereign RoadChain-Full: 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
2026-03-14 20:10:58 -05:00
parent 8ddac49849
commit 30967e4410
4 changed files with 912 additions and 0 deletions
--- a/ads/social/dev-to-articles.md
+++ b/ads/social/dev-to-articles.md
@@ -0,0 +1,303 @@
+# 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:**
+```markdown
+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:
+
+```bash
+sudo apt update && sudo apt upgrade -y
+```
+
+## Install Ollama
+
+One command:
+
+```bash
+curl -fsSL https://ollama.com/install.sh | sh
+```
+
+Verify:
+
+```bash
+ollama --version
+```
+
+## Load Models
+
+Pull the models you want. Here's what I run:
+
+```bash
+ollama pull llama3
+ollama pull mistral
+ollama pull codellama
+ollama pull phi3
+ollama pull gemma:2b
+ollama pull nomic-embed-text
+```
+
+List loaded models:
+
+```bash
+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:
+
+```bash
+sudo apt install -y hailo-all
+```
+
+Verify the accelerator is detected:
+
+```bash
+hailortcli fw-control identify
+```
+
+You should see: `Board Name: Hailo-8` and `26 TOPS`.
+
+## Test Inference
+
+Run a quick test:
+
+```bash
+ollama run llama3 "What is edge AI inference?"
+```
+
+For the Hailo-8 specifically, you can run object detection and classification models natively:
+
+```bash
+hailortcli run <model.hef>
+```
+
+## Serving Over the Network
+
+Ollama serves on port 11434 by default. To expose it to other nodes on your network:
+
+```bash
+OLLAMA_HOST=0.0.0.0:11434 ollama serve
+```
+
+Now any machine on your LAN can query:
+
+```bash
+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
--- a/ads/social/linkedin-posts.md
+++ b/ads/social/linkedin-posts.md
@@ -0,0 +1,276 @@
+# BlackRoad LinkedIn Posts
+
+**Principle:** Authority + Value-Expressive + Central Route (LinkedIn audience processes deeply)
+**Tone:** Professional but not corporate. Founder voice. Data-driven.
+
+---
+
+## Post 1: The Founder Story
+
+```
+I left finance to build AI infrastructure on Raspberry Pis.
+
+Not because I couldn't get a cloud budget.
+Because I did the math.
+
+One NVIDIA H100 on AWS: $3.90/hour.
+Run it 24/7 for a year: $33,696.
+For one GPU.
+
+Two Hailo-8 accelerators: $198 total.
+52 trillion operations per second.
+Run them forever. No bill. No vendor.
+
+I have a Series 7. A Series 24. A Series 65. A Series 66.
+
+I passed the same licensing exams Wall Street requires — then used that financial rigor to ask a simple question: "Why are we renting compute we could own?"
+
+The answer is that nobody told small teams they could own it. The cloud providers certainly didn't.
+
+So I built BlackRoad OS: self-hosted AI that runs on $400 of Raspberry Pi hardware. Five nodes. Sixteen models. Thirty websites. A billing system. A git server. Fifty AI skills.
+
+Monthly cloud bill: $0.
+
+The self-hosted cloud market hit $18.48 billion in 2025 (Grand View Research). Edge AI is growing at 21.7% CAGR.
+
+This isn't contrarian. This is where the math points.
+
+#SelfHosted #EdgeAI #AI #Infrastructure #Founder
+```
+
+---
+
+## Post 2: The Psychology Angle
+
+```
+I studied the Psychology of Advertising at the University of Minnesota (JOUR 4251, Dr. Claire Segijn).
+
+One stat changed how I build marketing:
+
+80% of advertisements are misunderstood by their audience.
+
+Not rejected. Not ignored. Misunderstood.
+
+The audience processes the ad and walks away believing something the advertiser never intended.
+
+Four mechanisms cause this:
+1. Omitted comparisons — "The best AI platform" (better than what?)
+2. Pragmatic inference — "May be the best" (may also be the worst)
+3. Juxtaposition — "Smart people choose X" (implies causation)
+4. Affirmation of consequent — "If you want Y, you need X" (false logic)
+
+So we made a rule at BlackRoad:
+
+Every claim must be literally, specifically, and verifiably true.
+
+"52 TOPS of neural inference" — true, measured.
+"16 Ollama models on a Pi 5" — true, run `ollama list`.
+"$0/month cloud bill" — true, there is no cloud.
+
+We even killed our own "30K agents" marketing copy when we realized it was aspirational, not factual.
+
+The truth effect (Fennis & Stroebe) says: the more people see a claim, the more true it seems. This works on lies AND truths.
+
+So we repeat truths. Obsessively. With sources.
+
+It's slower than hype marketing. It's also why our users stay.
+
+#MarketingPsychology #Advertising #Transparency #AI
+```
+
+---
+
+## Post 3: The Market Thesis
+
+```
+Three markets are converging on the same conclusion:
+
+1. Self-hosted cloud: $18.48B in 2025, growing 11.9% CAGR → $49.67B by 2034
+2. Edge AI: $24.91B in 2025, growing 21.7% CAGR → $118.69B by 2033
+3. AI inference: $106.15B in 2025, growing 19.2% CAGR → $254.98B by 2030
+
+The conclusion: inference belongs on the edge.
+
+Training requires cloud-scale compute. Always will.
+
+But inference — running models, answering queries, classifying data, powering agents — runs cheaper, faster, and more privately on hardware you own.
+
+A Hailo-8 accelerator costs $99 and delivers 26 TOPS.
+A cloud GPU costs $3.90/hour — you exceed the Hailo's price in 26 hours.
+
+94% of IT leaders fear vendor lock-in (Parallels 2026). 42% are moving workloads back on-premises.
+
+This isn't a prediction. This is happening right now, backed by $150B+ in market activity.
+
+BlackRoad OS sits at the intersection: self-hosted AI inference on commodity edge hardware.
+
+The shift doesn't require faith. It requires arithmetic.
+
+Sources: Grand View Research, MarketsandMarkets, Parallels
+
+#EdgeAI #SelfHosted #Infrastructure #AI #MarketAnalysis
+```
+
+---
+
+## Post 4: The Hiring/Culture Post
+
+```
+Our entire AI infrastructure runs on 5 Raspberry Pis.
+
+Here's what that says about how we build:
+
+1. We solve problems with architecture, not budget.
+A $55 computer is a datacenter node if you know Docker, WireGuard, and DNS.
+
+2. We own everything.
+Git server. Billing system. AI models. DNS filtering. Auth. Search.
+Nothing is rented. Nothing can be revoked.
+
+3. We measure in watts, not invoices.
+46 watts total. That's less than a light bulb.
+Not because we're cheap. Because efficiency is a design value.
+
+4. We verify before we claim.
+Every stat in our marketing is sourced. Every number is measured.
+We killed our own copy when it wasn't accurate.
+
+5. We document everything.
+207 repos on Gitea. Every decision, every config, every architecture choice.
+If I get hit by a bus, the system runs itself.
+
+This is the culture that builds BlackRoad OS.
+
+We're not hiring right now. But when we do, this is what we look for:
+People who'd rather own $400 of hardware than rent $33,696 of someone else's.
+
+#Culture #Startup #Engineering #AI #SelfHosted
+```
+
+---
+
+## Post 5: The Contrarian Take
+
+```
+Unpopular opinion: most AI startups are paying 100x too much for inference.
+
+Not training. Inference.
+
+Training a foundation model requires H100 clusters, thousands of GPUs, millions of dollars. Fair.
+
+But RUNNING a model? Answering a query? Classifying a document? Embedding text for search?
+
+That runs on a $99 accelerator plugged into a $55 computer.
+
+The AI inference market is $106 billion. Most of it is cloud inference — metered by the hour, billed by the token, scaled by the credit card.
+
+But inference doesn't need the cloud. Inference needs:
+- A model (free, open-source, download it)
+- Hardware (Hailo-8: $99, Pi 5: $55)
+- Software (Ollama: free, one command)
+
+Total: $154. Runs forever.
+
+The cloud GPU business model depends on you not knowing this.
+
+AWS doesn't want you to run `ollama serve` on a Raspberry Pi. That's a $33,696/year customer they lose.
+
+The edge AI market is growing at 21.7% because the secret is getting out.
+
+BlackRoad OS is how we prove it works — 16 models, 50 skills, 30 websites, $0/month.
+
+Not a pitch. A proof.
+
+#AI #Inference #EdgeComputing #SelfHosted #Startups
+```
+
+---
+
+## Post 6: The Data Sovereignty Angle
+
+```
+Your AI vendor's privacy policy is not a technical guarantee.
+
+It's a promise. Promises change.
+
+Technical guarantees look like this:
+- Pi-hole DNS filtering blocks 120+ tracking domains at the network level
+- WireGuard encrypts all inter-node traffic
+- No port is open to the internet (Cloudflare Tunnels for ingress only)
+- Models run on hardware in your physical possession
+- Data never transits a third-party network
+
+"Never leaves your network" is a physics statement, not a policy statement.
+
+The difference matters when:
+- A vendor changes their ToS (they will)
+- A government requests data (they can)
+- A breach exposes your prompts (it happens)
+- An acquirer gets your usage data (it's an asset)
+
+Self-hosted AI is not about distrust. It's about architecture.
+
+The same way HTTPS doesn't mean "I don't trust the internet" — it means "I'm not relying on trust when I can use encryption."
+
+Self-hosted means: I'm not relying on policy when I can use physics.
+
+#DataSovereignty #Privacy #AI #Security #SelfHosted
+```
+
+---
+
+## Post 7: The Education Flex
+
+```
+Most marketing teams optimize for clicks.
+
+We optimize for accurate comprehension.
+
+Sounds the same. It's not.
+
+Optimizing for clicks means:
+- Sensational headlines
+- Vague claims that imply more than they state
+- CTAs designed to create urgency
+- Metrics: impressions, CTR, conversions
+
+Optimizing for comprehension means:
+- Headlines with verified stats and named sources
+- Claims that are literally, specifically true
+- CTAs that reduce uncertainty instead of creating urgency
+- Metrics: reply rate, deploy rate, retention
+
+Why? Because the psychology (ELM, Fennis & Stroebe) says:
+
+Central-route persuasion — where the audience thinks carefully — produces attitude change that is DEEP and LASTING.
+
+Peripheral-route persuasion — where the audience uses shortcuts — produces change that is REAL but TEMPORARY.
+
+Clicks are peripheral. Deploys are central.
+
+We'd rather have 100 users who understand what BlackRoad does than 10,000 who clicked a flashy ad and bounced.
+
+That's not idealism. It's customer acquisition cost math.
+
+#Marketing #Psychology #ContentStrategy #AI #Startup
+```
+
+---
+
+## Posting Schedule
+
+| Day | Post Type | Frequency |
+|-----|-----------|-----------|
+| Monday | Market data / thesis | Weekly |
+| Wednesday | Technical / architecture | Weekly |
+| Friday | Culture / founder story | Biweekly |
+| Varies | Contrarian take | Monthly |
+| Varies | Psychology / education | Monthly |
+
+## Engagement Rules
+
+- Reply to every comment within 4 hours
+- Like every comment (signal that you're paying attention)
+- Never argue — redirect with data
+- If someone asks a question you can't answer honestly, say "I don't know" (builds more credibility than guessing)
+- Cross-reference blog posts when relevant (drives owned traffic)