active
Local LLM workbench
A repeatable rig for loading, quantising, and comparing open-weight models on a single 12 GB card — and finding the point where a smaller model stops being worth the memory it saves.
- llama.cpp
- GGUF
- Qwen
- PowerShell
Projects
Things that were actually built and run. Each page carries the setup, the measurements, and the parts that did not work.
Project list
ongoing
Agent guardrails
A tool-using agent with real permissions in a throwaway tenant, and a steady effort to talk it into doing something it should not. Notes on which controls held and which were theatre.
- Agents
- Prompt injection
- Graph API
- Entra ID
active
Detection bench
A small harness for writing KQL detections against simulated activity, so a rule's false-positive rate is known before it reaches anything that pages a human.
- KQL
- Sentinel
- Defender
- Atomic Red Team
More is on the bench
Identity lifecycle automation and a local retrieval pipeline are both in progress. They land here once there is something worth showing.