Now / Writing

Retrieval quality for the Headstarter semantic-search work can be measured with ranked results and usage signals instead of anecdote.

YOLOv8 detections plus embeddings over footage metadata can return timestamped, ranked investigation results faster than linear review.

KQL schema checks and cross-source correlation catch malformed non-native telemetry before it lands in detection tables.

Coursework in Machine Learning and Deep Learning can be turned into a repeatable comparison of model families on a fixed evaluation set.

Planned lab: test whether a trained retrieval or vision model can forget a specified subset without a full retrain.

Planned lab: simulate attacker/defender agents against the kinds of telemetry pipelines used in the Shaw Sentinel work.