Off-the-shelf APIs are fast to start, but many hospitality and enterprise teams need stronger data residency, lower unit cost at scale, or domain behavior that public models do not reliably provide. Llama-class open weights are a common base for a custom LLM stack.
“Custom” does not always mean full fine-tuning. Often the winning system is Llama serving + RAG + tools + policy, with optional adapters for tone or task formats.
Serving: vLLM/TGI or equivalent behind an internal AI gateway.
Knowledge: RAG over approved corpora with ACLs and citation checks.
Tools: MCP servers for read-first access to PMS/support systems.
Safety: input/output filters, PII redaction, and evaluation suites.
Hybrid: route sensitive prompts to private Llama; use cloud models for general tasks.
Use RAG when facts change often (policies, property content, runbooks).
Use LoRA/adapters when you need stable style, classification formats, or domain phrasing.
Reserve full fine-tunes for clear ROI after retrieval and prompting are exhausted—and keep a rollback path.
Version models and prompts, track quality regressions, set latency/cost SLOs, rehearse incident playbooks, and keep PCI/GDPR/SOC 2 boundaries intact when guest or payment context appears.
For a deeper serving walkthrough, see Deploy Llama 3 On-Premises on nitinrana.com.