Posts

I built a benchmark gate to decide a model upgrade. Its most valuable output wasn’t a score — it was discovering the upgrade couldn’t be run fairly on the hardware at all. A confident ’no’ is a first-class result.

An inference server that read as fully saturated was serving a fraction of its capacity. Both GPUs sat at 100 percent utilization while the real bottleneck was a single scheduler default set to 1.

A loop-engineered wireless-auditing agent, built as a controlled study in bounding autonomous offensive tooling. Its real subject is safe agentic automation: authorization enforced in code, not in a prompt.

The idea started with a problem I kept running into: it was late, I was done working in the lab, and I had no good way to shut everything down without walking downstairs or SSH-ing into three separate machines. The fleet was just going to sit there spinning, pulling power, until I got to it in the morning. The obvious answer was a shutdown script with a timer. But once I had the timer working, the next question was obvious too: if the bot can tell me to go to bed, why can’t it also put the fleet to bed on command? ...

The PyTorch Lightning supply chain incident triggered a reasonable response across the self-hosted AI community: audit your pip dependencies, verify your package hashes, check that pytorch-lightning is not in your environment. That is the right reflex, but it’s also incomplete. The dependency audit treats software packages as the threat surface. For AI infrastructure, the models themselves are part of the supply chain. A GGUF file downloaded from HuggingFace and loaded into Ollama or vLLM is not a static artifact: it contains executable code in its metadata, it shapes every inference call, and it has a provenance chain that most operators have never examined. ...

On April 30, 2026, two versions of the lightning package appeared on PyPI. Both were malicious. Socket’s automated scanner flagged them 18 minutes after publication. The maintainers quarantined them 24 minutes after that. Total window of exposure: 42 minutes. That sounds like a win. And for the detection side, it was. But in those 42 minutes, anyone who ran pip install lightning or whose CI pipeline did pulled down a hidden _runtime directory, a Bun JavaScript downloader, and an 11 MB obfuscated payload that fired automatically on import. The payload targeted environment files, API keys, GitHub tokens, browser credentials, and cloud provider secrets. Attributed to the TeamPCP group and connected to the broader Mini Shai-Hulud campaign, this was a coordinated supply chain attack against the AI engineering ecosystem. ...

There’s a moment in every homelab build where the pieces stop being parts and start being infrastructure. For VALKYRIE, that moment came when I ran a curl against a locally-served 27B parameter model and got a coherent response back using hardware I own, on silicon I paid $1,700 for, with no cloud provider anywhere in the chain. This is that build log. The Problem With Consumer GPUs The WWWOMBAT stack already had a working inference node: ODIN, an ASRock bare-metal Proxmox host running a VM with an RTX 3080 Ti passed through. Twelve gigabytes of VRAM. Enough for smaller quantized models, useful for daily driver work, but a hard ceiling for anything serious. ...

I wanted an IDS that would actually get my attention when something worth attention happened. Here is how I built one on a Raspberry Pi with a Suricata sensor, a deduplication engine, an LLM triage layer, and a Slack delivery pipeline.

The ATT&CK v14+ STIX data model quietly dropped the flat detection and data source fields every older tool relied on. Here’s what broke, why it was silent, and how the detection-strategy graph fixes it.

I built an entire website through a browser-based AI conversation. Then I switched to a terminal-first agent. Here’s what each approach actually gets right, and where each one falls apart.