Design decisions
Sharing One GPU Across Multiple Services →
How the Compute Foundation and Hardware Work
July 15, 2026← Back to ProjectsDesign and validation records ↓
Recorded status · Continuously running
Home Server Platform
Design, deployment, and operation of an information platform independent of external cloud providers.
Overview
Build and operate a practical information platform using only home computing resources, without depending on external cloud providers. Design, implementation, and operational monitoring are all handled solo.
Design and validation records
Measurements and validation
Confirming a Backup Actually Works →
How Monitoring and Automation Work
August 5, 2026Failures and lessons
What a Memory-Pressure Incident Taught Me →
How the Compute Foundation and Hardware Work
July 15, 2026Architecture
SYSTEMFigure 1 — Requests pass only through the encrypted path and reach services isolated at the container level.
Recorded capabilities
5 capabilitiesPhotos, files, and passwords can now be managed securely on a single personal PC, with no monthly subscription fees.
Can now securely reach the home PC from outside through an encrypted path, even while away.
Home appliances can now be controlled by voice.
Server anomalies (low disk space, a stopped service, etc.) are now detected automatically, with notifications sent out.
New services can now be added quickly without breaking the existing setup.
Technology Stack
Runtime
WSL2DockerDocker Compose
Reverse Proxy / Remote Access
CaddyTailscale
Photos / File Management
ImmichNextcloud
Auth / Password Management
Vaultwarden
Media Streaming
Jellyfin
Home Automation
Home AssistantMosquitto (MQTT)
Local AI
OllamaOpen WebUI
Monitoring / Operations
NetdatantfyWatchtower
Equipment
3 itemsEquipment actually used in this project, with its role and constraints. Each entry is based on a research-log record. Prices and availability are not listed.
Intel Core i7-12700
Used in this buildIntel
- Role
- CPU for every service. It is an everyday desktop PC whose resources are shared, not a dedicated server.
- Limitations & notes
- With everything on one machine, stopping this PC stops every service (single point of failure).
Record: How the Compute Foundation and Hardware Work Manufacturer page
NVIDIA GeForce RTX 3060 Ti
Used in this buildNVIDIA
- Role
- GPU shared by Jellyfin hardware transcoding, Immich machine-learning inference, Ollama local LLMs and ComfyUI image generation.
- Limitations & notes
- Because GPU memory is shared, ComfyUI is started only when needed. The card's board partner is not recorded.
Record: How the Compute Foundation and Hardware Work Manufacturer page
16 GB system memory
Used in this build- Role
- Memory shared by all containers on WSL2 and a robot-development workspace.
- Limitations & notes
- The practical bottleneck. A swap-usage alert above 98% led to memory limits on every container. An upgrade to 32 GB has not been done. Module part numbers are not recorded.
Roadmap
Phase 1
Foundation
Set up the virtualization layer and container platform, and established a secure remote access path. Laid the groundwork every later service runs on.
Phase 2
Data Migration
Gradually moved photos, files, and credentials that had been scattered across external cloud services onto the self-hosted platform, verifying integrity at each step.
Phase 3
Security Hardening
Audited the entire exposed surface, identified unnecessary exposure, and redesigned around the principle of least privilege, building in an ongoing audit process.
Phase 4
Extensibility
Introduced a reverse proxy and a monitoring platform, restructuring the setup so new services can be added quickly and with low risk.
Ongoing
Ongoing Operations
Continuing to expand into smart-home territory alongside day-to-day monitoring and maintenance, growing the platform through repeated cycles of planning and implementation.
Research log
7 articlesJuly 15, 2026
How the Compute Foundation and Hardware Work
The Docker-on-WSL2 runtime that runs the whole service fleet, and the hardware setup behind it, GPU included.
July 30, 2026How Access Control and Reverse Proxy Design Work
A design that only allows external access through an encrypted, dedicated tunnel, and centralized exposure management via a reverse proxy. Specific port numbers and IPs are intentionally omitted.
July 24, 2026How Photo, File, and Credential Management Work
Photo management (Immich), files/contacts/calendar (Nextcloud), password management (Vaultwarden). How data once left in external clouds was moved onto a self-hosted platform.
August 12, 2026How the Media Streaming Platform Works
The Jellyfin media server setup, and GPU hardware transcoding.
August 18, 2026How Home Automation and Voice Integration Work
How Home Assistant and Mosquitto (MQTT) sit at the center of a setup connecting Nature Remo appliance control and Alexa voice integration to a self-hosted platform.
August 24, 2026How Local AI and Private Search Work
Ollama running directly on the GPU, its Open WebUI front end, and SearXNG, a private search engine that never forwards queries externally.
August 5, 2026How Monitoring and Automation Work
Monitoring via Netdata, ntfy, and Watchtower, plus custom scripts for backups, health checks, and update tracking.