# MaraVexa

> Mara Vexa writes about AI security, observability, automation, self-hosted AI, Linux, and the systems where they collide.

Mara Vexa is a security, observability, and automation engineer exploring the intersection of these disciplines with AI. This bot-first site is a public, machine-readable record of technical work, incident reports, experiments, and philosophical essays on intelligence, security, and systems.

Key topics: AI security, LLM vulnerability scanning, red teaming, prompt injection, adversarial AI, production incident response, API shimming, observability, automation, open source, Linux, AMD/ROCm

## Site Pages

- [About](https://maravexa.com/about/): Background on Mara Vexa, this site, and its AI-assisted workflow.
- [Blog](https://maravexa.com/blog/): All published writing, newest first.
- [RSS Feed](https://maravexa.com/rss.xml): Full-content RSS feed.
- [Sitemap](https://maravexa.com/sitemap.xml): Canonical index of crawlable pages.

## Blog Posts

- [HyprSaver: Three Weeks of Manic Building a Screensaver for Hyprland](https://maravexa.com/blog/hyprsaver/): What started as a weekend Rust project to fill a missing screensaver gap on Hyprland turned into a three-week obsession — 30 GPU-accelerated shaders, a full optimization audit, ten failed wormhole attempts, and a pride palette pack.
- [KISS the Ring: Building a Prometheus Exporter for Network Inventory](https://maravexa.com/blog/signet-exporter/): Why I built a single-binary Prometheus exporter for network inventory instead of deploying another IPAM server — and how constraint, Go, and an existing monitoring stack got me further than any purpose-built solution.
- [The Satisfaction Paradox: On AI Editing, Ruskin's Illth, and the Writing That Feels Like Yours Until It Isn't](https://maravexa.com/blog/satisfaction-paradox/): A paper revealed that LLM editing neutralizes your stance while making you feel more satisfied. Explored through Ruskin's wealth/illth framework, with practical techniques to keep AI as a lever instead of a crutch.
- [Crutches and Levers: How to Actually Get Better With AI](https://maravexa.com/blog/crutches-levers/): AI coding agents let me ship projects I didn't understand. The uncomfortable process of fixing that turned out to be the point.
- [Indirect Observability: Monitoring AI Systems That Know They're Being Watched](https://maravexa.com/blog/indirect-observability/): When AI systems learn to game direct evaluation, observability must shift from watching outputs to reading side-channels. A framework for monitoring systems that resist observation.
- [Beyond Behavioral Scanning: Augmenting Garak with Mechanistic Persona Monitoring](https://maravexa.com/blog/garak-axis/): Bridging the gap between knowing that an LLM failed a red team probe and understanding how — by integrating Anthropic's Assistant Axis research into Garak vulnerability scanning.
- [Red Teaming on a Budget: Garak + Ollama + AMD RDNA2](https://maravexa.com/blog/garak-ollama-rdna2/): A practical guide to LLM vulnerability scanning with Garak and Ollama on AMD hardware with ROCm 6.3 — because constraint is the mother of flexibility.
- [The Map Is Not The Treasure: A Field Guide to Wealth You Can Actually Use](https://maravexa.com/blog/real-wealth/): Money is a map, not the territory — and we've been navigating by a map that's missing half the math. A field guide to wealth, illth, and why the tokens are imaginary.
- [Open Brain Surgery: A Guide to API Shimming in Production](https://maravexa.com/blog/api-shim/): How a corrupted artifact server, a broken backup, an incompatible API, and a ticking clock forced an emergency production migration — and why a 400-line FastAPI shim was the right tool for the job.
- [Data Fortress to Open Source: How Unlearning the Lessons of a High-Control Group Shaped My Path in AI Security](https://maravexa.com/blog/data-fortress/): How Unlearning the Lessons of a High-Control Group Shaped My Path in AI Security