When Accurate Entity Extraction Becomes an Infrastructure Problem
A retrospective on GLiNER, entity normalization, external verification, and why operational complexity led us to pause an NER pipeline.
Read article →Practical lessons on NLP, classification, clustering, backend systems, and turning messy streams of information into useful intelligence.
A retrospective on GLiNER, entity normalization, external verification, and why operational complexity led us to pause an NER pipeline.
Read article →A production audit of a rules-first XGBoost classifier reveals three false positives, a 98.25% sampling-weighted accuracy estimate, and the limits of model evaluation.
Read article →Why LLM publishing workflows need deterministic checks, bounded correction, and a separate authorization boundary.
Read article →Safe reprocessing, targeted story-cluster repair, and recovery-aware ML pipeline design.
Read article →Why database access patterns, bounded candidate sets, and sensible indexing matter as much as model speed.
Read article →Why AI-assisted development still depends on architecture, testing, and restraint—lessons from a production AI news platform.
Read article →Why small supervised models still matter alongside embeddings and LLMs—and when additional complexity is worth it.
Read article →A failed automation can have succeeded remotely. Here’s why blind retries create duplicate side effects—and how to design recovery more carefully.
Read article →Why embeddings are excellent for retrieving plausible matches but should not automatically decide whether two documents belong to the same real-world event.
Read article →Before embeddings, clustering, or LLM analysis, a domain-specific intelligence system has to decide which documents deserve to enter the pipeline.
Read article →