The lab is the proof of work behind everything else here. It is a bare-metal fleet, assembled largely from decommissioned enterprise hardware, that runs self-hosted AI inference, security tooling, and the infrastructure that supports them, entirely on equipment I own and operate. Nothing here rents a cloud GPU or leans on a managed service for the parts that matter. The hosts carry Norse names; the acronym the whole thing answers to is WWWOMBAT, the World Wide Web Orchestration of Models on Bare-metal Autonomous Topology.
Three GPUs, three tiers
The inference core is built on three generations of datacenter GPU, each matched to a tier of work rather than chased for novelty.
VALKYRIE is a Dell R740xd carrying two Tesla V100s, sixty-four gigabytes of HBM2 between them. This is the heavy tier: it serves a twenty-seven-billion-parameter model, Qwen3.6-27B quantized to AWQ, through vLLM, tensor-parallel across both cards. It runs on demand, waking when there is work and sleeping when there is not. A single one-line change to the inference scheduler more than tripled its sustained throughput, a story I tell in full on the blog.
LOKI is a large-memory Dell R640 with two Tesla T4s. This is the concurrent tier: agent orchestration, development work, and data-parallel benchmarking across the pair. The T4’s efficiency is the point, concurrent capability at a fraction of the V100 tier’s power draw.
TYR is a compact Dell R230 with a single Tesla P4, and the most unusual build of the three. The P4 is a passively cooled, slot-powered card designed for airflow-managed datacenter chassis; the R230 is not one. Running it here meant solving the thermals with ducting rather than a blower, and it holds temperature under sustained load. TYR is always on, and it hosts the services the rest of the fleet leans on.
A throughline across all three: every one of these is a passively cooled datacenter card running in a chassis that was never built to cool it. The thermal engineering, not the silicon, was the hard part.
Storage, backup, and sensing
ASGARD is a Dell R740xd running Proxmox Backup Server over a ZFS RAIDZ2 pool. Every host backs up into its own namespace, with cross-namespace deduplication, scheduled verification and pruning, and restores that are actually test-restored rather than assumed. It wakes for its backup window and sleeps afterward.
FREYR is a Dell R230 on an all-SSD array. It hosts the isolated wireless range that GUNGNIR runs against, and is being stood up as the fleet’s log-aggregation node.
HEIMDALL is a Raspberry Pi, always on, running a Suricata sensor that feeds the intrusion-detection pipeline.
Capabilities
Beyond the hardware, the fleet runs as an operated system.
- Backup and recovery. Every reachable guest has a recovery point in deduplicated storage; representative restores are actually test-restored, and scheduled verify and prune keep them honest.
- Power-aware orchestration. The heavy and backup tiers are on demand: they sleep on a nightly schedule and wake on command or automatically when a backup is due, with a guard that refuses to cut power to a host mid-backup. Full capacity when it is needed, near-idle draw when it is not.
- Self-hosted source and PKI. A self-hosted Forgejo forge keeps source off the public hosting surface, secured by wildcard TLS issued automatically across the internal domain, reachable off-network only through an identity-gated tunnel.
- Intrusion detection. The Tattle pipeline takes sensor events through deduplication and automated triage to alerting, with canary and honeytoken triggers seeded across the environment.
This is the ground the research and the consulting practice stand on. The agentic security work, the AI-governance tooling, and the writing all run here, on infrastructure I built and can account for end to end.