Add note: Practical product-name extraction with a compound pipeline

[herohua]

Summary

Adds a new tech-note distilling the compound NER + entity linking pipeline used to extract product-name mentions from documentation content. Written for software engineers and technical PMs without an NLP background.

The note covers:

• Why naive approaches (single regex, LLM-only) don't survive contact with real content
• A nine-stage cascade: dictionary spotter → exclusion masks → boundary expansion → dedupe → fuzzy match → common-word filter → LLM verifier → tagged emission
• Why the LLM runs last (with explicit comparison to "LLM first" and "LLM only" alternatives)
• Honest limits of the design
• When to use this shape and when to reach for a SOTA alternative instead
• How each stage maps to standard names in the entity-linking literature

Dated 2025-11-07 to reflect when the pattern was identified in the source projects.

Test plan

• Frontmatter renders correctly on the site (title, date, tags, publish flag)
• All inline citation links resolve
• ASCII pipeline diagram renders in a monospace block
• Tables render correctly (literature/SOTA tables)

🤖 Generated with Claude Code