13. Industry Applications

Chapter focus: The last mile from "the model works" to "the business works." This chapter covers both AI-compute side (knowledge augmentation RAG, LLM serving) and HPC side (AI surrogate models accelerating simulation, end-to-end scientific problem delivery) — the real-world closed loop of "AI × scientific computing."

Articles

1. Kubernetes + RAGFlow Quick-Deploy Guide — The "golden path" for single-node K8s + RAGFlow intelligent retrieval deployment, with mirror acceleration and persistence config.
2. Science Computing Embraces AI: Surrogate Models in HPC Workflows — Turning "hour-scale simulation" into "millisecond-scale prediction" with AI surrogates: HPC data pipeline, model selection, error control and re-verification loop.
3. Battery Lifetime Prediction: An End-to-End Case from Scientific Problem to Software Delivery — "One problem, one software" in action: problem definition → data → AI model → HPC validation → two-phase delivery (offline diagnosis / online early warning).
4. Enterprise Knowledge-Base Q&A, Production-Grade — Chunk strategies, hybrid retrieval, rerank, evaluation metrics and four productionization concerns — answering "why it answers wrong" and "how to keep it reliable."
5. Private LLM Serving Architecture: From GPU Selection to Production HA — Memory estimation and GPU selection, vLLM's three mechanisms, multi-replica/canary/failover architecture, load-test-driven capacity planning.
6. Scientific Agent Deployment: From Tool Chains to Autonomous Research Workflows — On a stable inference service, upgrade single-turn Q&A into multi-step research task execution: tool packaging, four reliability lines, semi-autonomous boundaries and task-set evaluation.

📌 Planned: RAG evaluation metrics deep-dive, agent workflow rollout assessment, industry-scenario ROI estimation, AIOps application retrospectives.

Reading Path

• Application deployment: start with the RAGFlow deploy to build a knowledge-base foundation, then use the production-grade article to reach production retrieval quality; for intranet LLM serving, read the private serving architecture article.
• AI × scientific computing: build architecture intuition from the surrogate-model article, walk the full loop with the battery-lifetime case, then automate the process with the scientific-agent article.
• Scenario evaluation: quantify your own rollout path using each article's evaluation/retrospective criteria.

Key Takeaways

• The foundation of knowledge-augmented applications is reproducible deployment: a documented golden path with environment/image/storage done in one pass.
• Retrieval-quality optimization order = chunk strategy → hybrid retrieval → rerank → embedding upgrade; decide each layer with eval numbers.
• The core of AI-accelerated scientific computing is surrogate + re-verification loop: HPC produces data, AI accelerates, HPC re-verifies — AI computing and HPC reinforce each other.
• The engineering core of application rollout is observability and cost control: the rule layer guarantees determinism, the model layer covers open-endedness.