Grab's engineering team shares how they evolved a single internal support bot into a company-wide AI agent platform called LLM-Kit, now powering over 500 services. The post details the specific pain points encountered — lack of evaluations, slow model switching, fragmented observability, and weeks of boilerplate setup — and how each became a core framework primitive. LLM-Kit provides a scaffolded FastAPI+LangGraph template with pre-wired OIDC auth, Vault-based secrets, OpenTelemetry tracing, MCP server integration, gRPC service connectivity, and built-in eval endpoints (ROUGE, BLEU, LLM-as-judge). What previously took two weeks of production wiring now takes about an hour. The post is Part 1 of a series; Part 2 will cover the GrabGPT Gateway, remote MCP framework, and evals platform.
