Systems Engineer | Homelab Enthusiast | AI Researcher | Autistic
Started my journey 13 years ago as a Cloud DevOps engineer. Over the years I've worked in security, project management, software engineering, and AI/ML engineering—now I kind of do it all. With one major exception: I'm doing everything in my power to push back against the cloud being the default for every task, no matter how large or small. Besides just being a nerd, this is why I got into homelabs.
This homelab replaces approximately $11500/month in cloud services. Over 3 years, that's $414,000 in potential savings after hardware costs.
All 23 VMs and their data remain on-premises. No cloud provider has access to databases, secrets, or application data.
Managing 7 hypervisors, 268 vCPUs, and various services provides hands-on experience with enterprise-grade infrastructure.
Full control over 575GB RAM and 52TB storage allocation. No vendor lock-in or arbitrary service limits.
Run bleeding-edge ML research on consumer GPUs—no cloud A100s needed. Custom Mojo kernels achieve 12x speedups over PyTorch.
Full control over model hidden states enables vulnerability research impossible on cloud APIs. Your data never leaves your network.
| Name | Node | Type | CPU | RAM | Disk | |
|---|---|---|---|---|---|---|
| mattermost | homelab2 | vm | 2 | 4G | 30G | |
| metasploitable2 | homelab2 | vm | 1 | 1G | 8G | |
| OpnSense | homelab2 | vm | 4 | 4G | 64G | |
| attack-box | homelab2 | lxc | 2 | 2G | 16G | |
| vuln-docker | homelab2 | lxc | 4 | 4G | 31G | |
| kali-netdash | homelab3 | vm | 8 | 32G | 80G | |
| TimeScale | homelab1 | vm | 16 | 64G | 256G | |
| OMV | homelab1 | vm | 4 | 8G | 32G | |
| Prefect | homelab1 | vm | 8 | 8G | 96G | |
| RabbitMQ | homelab1 | vm | 8 | 8G | 96G | |
| SubSeek | homelab1 | vm | 4 | 16G | 32G | |
| Docker | homelab1 | vm | 16 | 8G | 128G | |
| forgejo | homelab1 | vm | 4 | 16G | 64G | |
| PyServices | homelab1 | vm | 8 | 8G | 96G | |
| DataServices | homelab1 | vm | 8 | 16G | 128G | |
| admin | mercury | vm | 4 | 8G | 64G | |
| site1 | mercury | vm | 4 | 16G | 64G | |
| redis-server | mercury | vm | 5 | 8G | 32G | |
| API-Gateway | mercury | vm | 4 | 16G | 64G | |
| Caddy-Router | mercury | vm | 4 | 4G | 32G | |
| Temporal | mercury | vm | 4 | 16G | 32G | |
| s3 | mercury | vm | 4 | 8G | 32G | |
| PowerDNS | mercury | vm | 2 | 4G | 32G | |
| workstation | local | workstation | 128 | 256G | 3600G | |
| pop-os-c2 | homelab4 | vm | 8 | 32G | 500G | |
| nexus | homelab4 | vm | 4 | 8G | 40G |
lab-stack-backend
lab-stack-blog
code.forgejo.org/forgejo/runner
cloudflare/cloudflared
infisical/infisical
postgres
We're building an adaptive memory system for AI inference — memory that operates at the activation level, not the token level. Early Gemma-4-31B results: a 3.12 selectivity ratio for on-topic vs adversarial recall. Provisional patent filed.
Jane Street published the Dormant LLM Challenge answer key today. My March submission claimed all three triggers. The answer key disagrees on every model. Here's what I actually got wrong, the methodology error that drove it, and what I'd do differently with the benefit of hindsight.
Three things landed this week at Light of Baldr — an open SAE dataset for Gemma-4-31B, a working MAX Engine stack on DGX Spark, and early signal from an activation-level memory system.