Personal AI Platform

Mathias

Digital Product Manager & Technology Consultant  ·  Sweden

A self-hosted AI development environment — from bare metal to running agents. Built for privacy, production quality, and speed of iteration.

All systems operational Open source tooling Gitea-native No cloud lock-in
What it enables

Six capabilities.
One platform.

01
Private AI Inference
GPU-accelerated local LLM serving with 27+ model routes. Data never leaves the network. Cloud APIs available as optional tier.
02
Agent Orchestration
Executor + reviewer loops for autonomous coding tasks. Risk-tiered tool middleware. Flat peer-to-peer architecture, not hierarchical.
03
Financial Automation
Invoice extraction, validation, and ISO 20022 payment generation. Fortnox integration with full audit trail and multi-tenant isolation.
04
Knowledge Management
Self-hosted brain with BM25 + vector hybrid retrieval. Persistent memory across sessions, machines, and AI harnesses.
05
Private Git & CI/CD
Gitea as canonical remote. Flux GitOps, Gitea Actions, buildah OCI builds. Full code workflow without GitHub dependency.
06
Full-Stack Development
Go + HTMX services from local dev to production Kubernetes pods — same day. Templates, scaffolding, and one-command deploys.
How it's built

Hardware to protocol.

Hardware
koala
RTX 5070 · k3s cluster node · GPU inference · Tailscale hub
iguana
M2 Ultra Mac · Services · Builds · Ollama models
flamingo
Mac mini · Daily driver · Edge node · Claude Code
piguard
Raspberry Pi · Network gateway · LiteLLM proxy
piblock
Raspberry Pi · NAS · Restic off-host backup target
Platform
k3s + Flux
Kubernetes on koala · GitOps reconciliation · all manifests in git
Tailscale mesh
Zero-trust networking across all five machines · no VPN config
Observability
Prometheus · Grafana · Jaeger distributed tracing · Loki
Identity & Secrets
Dex OIDC · SOPS-encrypted manifests · k8s secretKeyRef
CI/CD
Gitea Actions · buildah OCI builds · local registry · Flux apply
AI Stack
LiteLLM proxy
27+ model routes · local + cloud · single unified API
llama-swap
GPU model manager · hot-swap between slots · RTX 5070 offload
MCP protocol
brain · gitea · infra · routing — Dex-authenticated servers
Brain KB
BM25 + pgvector hybrid retrieval · LLM synthesis · 100+ entries
ADK agents
Go ADK · kimi-k2.6 executor · gemma4-31b reviewer · OTLP traces
By the numbers

What's running.

5
Machines on
Tailscale mesh
27+
LLM routes
local + cloud
4
MCP servers
in production
12GB
GPU VRAM for
local inference
100+
Brain knowledge
entries
1day
Idea to
production deploy