Entity-driven news intelligence app — built solo, production-running. Ingestion pipeline processes 49,800+ articles across a knowledge graph of 130,000+ entities. Redesigned article processing architecture to improve pipeline reliability from 71% to 95.7%. Cut AI inference costs by 75% (₹500–1,000/day → ₹100–250/day) by migrating 98% of extraction workload to DeepSeek with Gemini as automatic fallback. Key systems built from scratch: — Multi-model LLM orchestration with cost-aware routing — pgvector semantic deduplication with entity-overlap + headline Jaccard fallback — Non-destructive identity resolution layer (canonical entity grouping across 130K+ entities without rewriting foreign keys) — Alias contamination detection to reject LLM-injected co-reference noise