We are moving away from LLM-based industry classification in favor of a fast, deterministic prototype-per-label embedding approach. This new architecture, detailed in updated documentation, ensures better consistency and lower latency by utilizing offline vector similarity while maintaining integration with existing rules and pipelines. The system is designed to provide high-recall classification while keeping resource usage low for local deployments.

Added a new resume-matching module that computes a resume_similarity_score for every job by comparing skill coverage against a configured taxonomy. The implementation supports both direct LLM extraction and a fully deterministic offline fallback, ensuring the pipeline remains robust without mandatory external dependencies. This feature is now integrated into the final scoring pipeline, meaning resume-to-job relevance now directly impacts job ranking.

This update introduces a resume-matching feature to the job filtering pipeline, replacing the previous neutral fallback score for resume similarity. The pipeline now dynamically compares job descriptions against a provided resume and a new skills taxonomy, weighting resume compatibility alongside seniority, location, and industry relevance. This ensures that the final job scores are better aligned with the user's professional background.

The Jobinator project has been initialized, providing a minimal pipeline to source, filter, and score job opportunities based on user-defined criteria. The pipeline orchestrates scraping via JobSpy, data merging, specialized filtration including seniority and industry mapping, and atomic exports to a structured Excel workbook. This foundational code is ready for automated, human-assisted job searching and prioritization.