Our first reputation risk classifier was precise but missed about half the risks. Our second version changed the training data, the labels and the model, and raised recall by roughly 0.3 while improving precision too.
Featured story
Searching Journalists by Topic with LLM-Generated Fluid Topics and Embeddings
How we replaced keyword matching on articles with semantic search over short, LLM-generated topic labels: curation, embeddings, two-stage kNN on OpenSearch, and the parts we had to rework along the way. Read the story