Family
Föräldraledighetsplaneraren
Agent-ready leave planning for Swedish families
Complete days, pay and trade-offs from phone or assistant — shareable plan, organic SEO wedge.
CTA
Complete plan
AI venture studio · Decision-first portfolio. Measured learning. Agents inside policy — never as fake executives.
Operator
Principal
Strategy, values and irreversible decisions — approve on phone, not in meetings.
Learning loop
Signal
PostHog, inbox, market
Decision
Expected outcome set
Work
Bounded execution
Outcome
Measured vs expected
Lesson
Correction → memory
Projects
Live projects
Focus projects run measured experiments. Monitor projects stay visible with honest status. Parked work does not pretend to be active.
Focus
Active validation and daily growth attention — platform plus product bets.
Monitor
Live and visible, outside the daily growth queue.
Parked
Dormant — not shown as Active.
Family
Agent-ready leave planning for Swedish families
Complete days, pay and trade-offs from phone or assistant — shareable plan, organic SEO wedge.
CTA
Complete plan
Sports
Club intelligence digestible by coaches and agents
Proactive hockey analyses before dashboards — built for mobile review and assistant handoff.
Offer
Club Pack beta
Studio
Field Notes & Company OS distribution
Evidence-based notes for operators building AI-native companies — decision-first, not dashboard-first.
CTA
Subscribe
Studio
Mobile control plane for portfolio decisions
Signal → decision → work → outcome → lesson. Agents inside policy; human gates on consequential actions.
Mode
Platform
Studio
Tool permissions & approval levels for agents
Governance layer so assistants and workflows cannot exceed policy — audit every consequential call.
Mode
Monitor
Sports
Consulting surface adjacent to THA
On-demand analytics for teams without a full-time hire — Monitor while THA Club Pack validates.
Mode
Monitor
Hyperlocal GTM from mobile + assistant
Physical mail and referral loops for local sellers — pilot validates agent-assisted campaign creation.
Mode
Monitor
SaaS
Enterprise agent infrastructure narrative
Content surface on distributed agents and sovereign AI — Monitor until a wedge product appears.
Mode
Monitor
SaaS
Workflow cards operators can hand to agents
Describe a workflow, get a bounded automation with handover docs — agent-native delivery pattern.
Mode
Monitor
Studio
Personal trust realm — distribution only
Founder portfolio in Personal realm; linked here as distribution, not a studio venture.
Realm
Personal
Studio
Problem-first signal ingestion for the studio
Collect trends, competitor moves, inbox friction and portfolio metrics; rank at most three decisions per day for the operator.
Output
Daily signals
Studio
Assistant-connectable Company OS surface
Scoped read/write for agents: portfolio state, pending approvals, validation contracts — every tool call audited under MeshGuard policy.
Mode
Concept
Powered by TUR Company OS
TUR Company OS is the control plane for this studio — portfolio hierarchy, approval gates, event memory and measured learning loops. Agents draft and execute inside proven workflows; humans keep governance at consequential boundaries.
10+
Live ventures
4
Focus projects
Signal → Lesson
Learning loop
Decision-first
Every action traces to a signal, decision, expected outcome and human correction — not a dashboard metric wall.
Focus · Monitor · Parked
Focus projects run measured experiments. Monitor projects stay visible. Parked work does not compete for daily attention.
Human approval gates
Publish, deploy and external send never run autonomously. Mobile approvals keep velocity without losing control.
The thesis
An AI-native venture studio where decisions, experiments and outcomes compound in Company OS — agents execute proven workflows under human approval gates.
Evidence before automation
Signal → decision → work → outcome → human correction. Agents join only after the workflow is stable enough to evaluate.
Focus projects in rotation
Company OS, THA, Parental and TUR Field Notes run measured experiments. Everything else is Monitor or Parked.
Organic visibility in the AI era
Canonical claim pages, answer-engine FAQ, Field Notes and product tools — not autonomous blog spray across every domain.
Daily pipeline
Ingest signals, draft under MethodPolicy, approve on phone, publish only with a gate. Self-hosted jobs still power research and IndexNow where useful.
PostHog Focus funnels, Raindrop/Gmail learning labels and Tech Pulse digest surface at most three mobile decisions.
PostHog · Company OS · Resend
Each Focus project keeps one active validation contract. Distribution packs stay bounded to one claim surface + one redistribution asset.
Validation contracts · Field Notes
Marketing and development drafts run free-first models under spend gates. Deterministic policy still owns approvals and budgets.
OpenRouter · Inngest · MethodPolicy
Linear Mobile / operator approvals. No autonomous publish, deploy or external send from an agent.
Approve · Defer · Reject
Approved drafts can commit, deploy and IndexNow after Level 2/3 policy. Outcomes return to Growth Memory.
GitHub · Vercel/Netlify · IndexNow
Stack
Every OS phase has named tools so partners can see exactly how we sense, compose, deploy, and audit autonomous ventures.
Everything runs on one Hetzner VPS managed by Coolify. No managed databases, no SaaS overhead.
Coolify
Self-hosted PaaS that manages all containers, env vars, and deploys on the Hetzner VPS.
Hetzner
CX32 VPS in Frankfurt. ~150 kr/month. Runs everything — n8n, LiteLLM, Minio, Umami, OpenClaw, Qdrant.
Minio
S3-compatible object storage for draft files, keyword JSONs, and performance insight data.
n8n runs 23 active workflows. Every schedule, API call, conditional branch, and Telegram notification.
n8n
Self-hosted workflow engine. Handles keyword research, content generation, publishing, analytics, and error monitoring.
Telegram
All notifications and approvals arrive here. Inline buttons for approving EIK articles without opening a browser.
OpenClaw
Webhook publish server. Receives article payloads and commits markdown to GitHub — one HTTP call per publish.
Paperclip
Governed agent workflows with spend gates. Strategy and cost reviews surface as signals — not autonomous executives.
LiteLLM proxies all model calls. Gemini 2.5 Flash runs article generation for 6 sites. Groq handles keyword scoring. Costs tracked per site.
LiteLLM
Self-hosted LLM gateway. Routes to Groq, Gemini, or other providers. Tracks spend. Single endpoint for all models.
Groq
Fast inference for keyword scoring and EIK article generation. Llama models. Free tier covers daily usage.
Gemini
Gemini 2.5 Flash is the primary article generator for 6 sites — TUR, TAN, TAF, TPR, THA, THB. Large context window, routed via LiteLLM.
Qdrant
Self-hosted vector DB for semantic search and RAG experiments on article archives.
PostHog tracks Focus funnels across studio and venture sites. Weekly review surfaces conversion and CTA evidence.
PostHog
Product analytics and funnel tracking. CTA clicks and venture card engagement — filter by site property.
IndexNow
Instant URL submission to Google and Bing after every article publish. No waiting for crawlers.
Listmonk
Self-hosted newsletter platform. Subscriber list synced with Resend for transactional delivery.
Numbers
Rough monthly cost to run Focus distribution, self-hosted jobs and the Studio site — not a claim of zero-headcount autonomy.
“Self-hosted jobs still run on one Hetzner VPS where they beat managed SaaS on cost. Company OS stays on Vercel + Supabase.”
“Model spend is gated by MethodPolicy and free-first routing. Drafts are cheap; publish still needs approval.”
“Focus domains first: theunnamedroads.com, thehockeyanalytics.com, tur-parentalleave — Monitor sites stay live but outside daily growth ranking.”
What this is
An AI venture studio — one founder, 10+ live projects, self-hosted AI infrastructure.
The Unnamed Roads is an AI venture studio — one person, 10+ live projects, and an AI agent pipeline that handles everything from keyword research to publishing.
The entire stack is self-hosted on a single Hetzner VPS managed by Coolify. n8n orchestrates ~23 workflows. LiteLLM routes LLM calls to Groq and Gemini. OpenClaw commits markdown to GitHub. Netlify and Vercel deploy. Total cost: ~400 kr/month.
This site is where I write about what I'm building, what breaks, and what it actually costs to run AI infrastructure at this scale. No fluff. Just systems.
Fully self-hosted
Coolify + Hetzner replaces 2 300+ kr/month in managed SaaS. One VPS runs everything.
Autonomous publishing
Content research, generation, and publishing run on cron schedules without manual steps.
Telegram-first ops
Every workflow completion sends a Telegram message. Approvals use inline buttons. No dashboards needed.
10+
Live projects
400 kr
Monthly infra
23
n8n workflows
Say hello
If you're building something similar — self-hosted AI infra, multi-site content ops, n8n workflows — I'm happy to talk. No sales pitch, just notes from the field.
Also on Bluesky: @theunnamedroads.bsky.social