Short and ambiguous queries are a common challenge in job search systems, often leading to poor matches and reduced user satisfaction. We present a production-ready, LLM-powered query reformulation framework that improves low-performing queries by generating richer, context-aware alternatives. To address off-topic reformulations and hallucinations often introduced by LLMs, our hybrid offline–online system incorporates domain knowledge, applies semantic filtering, and uses a weighted fusion ranking approach to improve retrieval effectiveness while maintaining low latency. Offline evaluation shows consistent 10% gains in NDCG@10. In a user study, 87.2% of reformulations were perceived to improve intent clarity, and 78.8% improved perceived search quality. Our interactive demo ( https://ecir26-demo.static-upwork.com ) offers an end-to-end walkthrough of the system.