What problem does it solve? Manually reviewing every scraped job listing is slow and inconsistent. This Skill implements a two-stage pipeline that filters irrelevant jobs locally with embeddings and then applies LLM-based deep scoring, so only high-match jobs reach the application or approval stage. ## Core Features & Use Cases - Two-Stage Filtering Pipeline: Stage 1 uses local Ollama embeddings (nomic-embed-text) with cosine similarity to discard low-relevance jobs; Stage 2 sends survivors to Groq Llama-3-70b for structured JSON scoring. - Structured Output Validation: A Pydantic schema enforces score (0-100), reasoning list, critical flags (visa sponsorship, language requirements, seniority), and a decision of auto_apply, ask_user, or reject. - Confidence Threshold Routing: Scores of 85+ route to AUTO_APPLY, 60-85 to AWAITING_APPROVAL via Telegram, and below 60 to DISCARDED, with results persisted to the job_listings table. - Use Case: After scraping 200 job postings, run the pipeline to discard 80% locally at zero API cost, then LLM-score the top 20% and automatically queue only the strongest matches for application. ## Quick Start Score the newly scraped job listings against my profile and route them by confidence threshold.
