← Back
Architecture in operation
The stack behind the work.
Two environments shaped by the same principle: data, applications, models, and operations should form one understandable system. The Unnamed Roads optimises for autonomy and low cost; the enterprise stack optimises for governed scale.
- live projects
- 10+
- containers
- 30+
- MCP integrations
- 12+
- monthly infrastructure
- ~€35
Emil Ingemar Karlsson — 2026 Agentic Engineering Stack
Paperclip-first cockpit · LangGraph 24/7 orchestration · capability-scoped agents (not org-chart). Emil steers kanban; LangGraph polls todo-issues, plans via LiteLLM, triggers agent heartbeats; nothing ships without approval-webhook + Telegram.
Interface — control towerWhere Emil creates work & approves output
kanban
Paperclip
paperclip.theunnamedroads.com
mobile
Telegram
push + inline approve
automation
n8n
n8n.theunnamedroads.com
dev
Cursor
local · human override
todo / backlog issues
Orchestration — LangGraph + Paperclip agents24/7 dev & data passes · systemd heartbeats as backup
LangGraph worker
Paperclip todo→LLM plan (LiteLLM)→comment + in_progress→agent heartbeat→in_review→Telegram notify
Developer Agent
Data Engineer
QA Agent
Copy Agent
CMO Agent
Langfuseevery LiteLLM call traced · plan steps · agent heartbeats · cost/latency
MCPJungle invoke · memory · sandbox · paperclip
Tools — MCP LayerMCPJungle gateway · mcp.theunnamedroads.com · routes to scoped tool servers
agentic-stack-tur (Coolify) — agentic-primitives via mcp-bridge
orchestrator
langgraph-worker
poll · plan · heartbeat · checkpoint pg
memory
memory-mcp
pgvector · 768d HNSW
sandbox
sandbox
python · node · bash
bridge
mcp-bridge
agentic-primitives MCP
db
pgvector
memory + langgraph checkpoints
External MCP servers via MCPJungle
paperclipgithubfilesystemfetchpostgres-umamipostgres-n8nresendcloudflarenetlifybitwardenclarityn8ninference · embeddings · all Langfuse-traced
ModelsLiteLLM — single endpoint · auto-fallback · redis cache · Langfuse callback
Claude Sonnet 4.6
Anthropic · Foundry
code · analysis
Claude Haiku 4.5
Anthropic
LangGraph plan steps
Gemini 2.5 Flash
Google · free
long ctx · vision
Cerebras Qwen3 235B
Cerebras · free
workhorse
Groq Llama 3.3 70B
Groq · free
latency-critical
DeepSeek V3 / R1
DeepSeek
$0.27/1M tok
Mistral Small
Mistral · EU
free tier
gemini-embedding-001
embeddings · 768d
used by memory-mcp
Governance — human-in-the-loopNothing merges, publishes, or deploys without explicit approval
host
approval-webhook
:9191 · systemd on tha
policy
No auto-publish
store-draft · content-review only
budget
LiteLLM alerts
paperclip-agents virtual key
all services on
Coolify — theunnamedroads-platformSame project as Crawlab · Umami · Paperclip · LiteLLM · n8n
agentic-stack-tur
langgraph · pgvector · sandbox · memory · mcp-bridge · langfuse
paperclip-tur
kanban cockpit
litellm-tur
model router + cache
openclaw-tur
drafts · publish scripts
MCPJungle
mcp.theunnamedroads.com
n8n · Umami · Crawlab
automation · analytics · scrape
runs on
InfrastructureHetzner VPS · Docker · Traefik TLS · Coolify orchestrated
Hetzner VPS
Helsinki · 8 vCPU · 16 GB
Coolify
coolify.theunnamedroads.com
Docker + Traefik
coolify network bridge
Postgres × N
Paperclip · LiteLLM · n8n · Umami · pgvector · Langfuse
ClickHouse + MinIO
Langfuse traces & events
systemd timers
paperclip heartbeats · budget alerts
Observability — cross-cuttingWraps every layer · ground-truth for cost, quality, regressions
tracing
Langfuse v3
langfuse.theunnamedroads.com · in agentic-stack-tur
notify
Telegram
approvals · agent-ready · daily spend
verify
Smoke scripts
verify-daily-cockpit · verify-langgraph-stack
agentic-stack-tur (Coolify)
Paperclip cockpit
MCP / tools
Governance / approval
Observability
Emil Ingemar Karlsson — Enterprise Delivery Stack
Enterprise flow for daily delivery: build in Cursor, enforce policies in a guardrail layer, execute on Databricks and Azure, and ship through controlled ADO pipelines.
Interface — daily engineering cockpitWhere work is authored and reviewed
dev
Cursor
agent-assisted coding · SQL · infra changes
planning
Backlog + PRs
ticketed changes · review-first workflow
policy + quality gates
Rules & Guardrail Layer — policy enforcementControls before anything reaches production
policy
Security & Naming Rules
patterns · conventions · ownership checks
validation
Data & Test Validation
schema checks · quality tests · smoke tests
approval
Approval Workflow
manual checkpoints for high-impact changes
approved delivery path
Execution & Runtime — enterprise platformsCore enterprise platforms
data platform
Databricks
pipelines · jobs · ML/AI workflows
cloud
Azure
storage · identity · governed services
release
Azure DevOps (ADO)
CI/CD pipelines · environments · release gates