Writing
Notes from production
Applied AI in production environments, backend architecture, and notes from doctoral research. Written to be useful, not to rank.
-
Catastrophic forgetting, measured
A small network lost 12.3 points of accuracy learning a second task, real but short of my registered threshold for severe. EWC recovered 98% of it.
-
Blackwell quantisation on my own card
Q4_K_M more than doubled throughput and cut VRAM by two thirds on my RTX 5070 Ti, staying 98% similar to full precision. The interesting part was where that 2% lived.
-
What the EU AI Act actually requires of a masking layer
The Act never says masking. Three articles, on data, logging and human oversight, each describe a property a masking layer either has or does not.
-
Your model is not the product. The pipeline around it is.
Swap the model and almost nothing changes. Swap the masking layer, the permissions, or the monitoring, and the system stops running or starts leaking.
-
Detecting PII before the model sees it
Two registered experiments, one frozen dataset, two detectors that each miss a different category. The entity-level numbers say which, and by how much.
-
Why I pre-register a kill-gate for every experiment
A kill-gate is the number, written down in advance, at which I abandon my own hypothesis. How three real gates fired, passed and refused to answer.
-
Permissions that travel with the person
An AI assistant should own no permissions. It borrows the rights of the person using it, and that single design choice bounds every attack path.
-
What an internal MCP server actually solves
An internal MCP server does not make the model smarter. It gives authentication, permissions and audit a place to live between assistants and data.
-
The demo is the easy part
A demo proves the happy path exists. The system that runs every day on uncurated inputs still has to be built, and that is most of the work.
-
The masking layer matters more than the model
A masking layer decides what sensitive data ever reaches a language model. Why that design choice outweighs model selection in production AI.