IECSA: Enterprise AI Agent
A production-oriented customer-success agent prototype that brings Spring backend discipline to LLM workflows.
The Goal
Customer-success requests are rarely one API call. The agent needs accurate policy context, controlled access to customer data, session isolation, and a safe way to escalate risky decisions.
IECSA (Intelligent Enterprise Customer Success Agent) is a hands-on Spring AI prototype exploring that full path: an LLM can reason over a request, retrieve grounded context, call the right business tools, and hand off decisions that require approval. It is deliberately presented as a prototype—not a production AI deployment.
Architecture
What I Built
- Agentic tool loop: Java
@Toolmethods and Spring AIChatClientfor customer-account and workflow actions. - MCP integration: a local JSON-RPC CRM server, batched customer lookup to avoid N+1 calls, and propagated identity context for row-level access control.
- Grounding: an in-memory vector store with local ONNX embeddings and
QuestionAnswerAdvisorto retrieve policy context before answering. - Memory and orchestration: windowed per-session memory, asynchronous long-term fact extraction, and supervisor-worker delegation with isolated worker contexts.
- Safety and readiness: human-in-the-loop approval flow, prompt-injection filtering, request lifecycle handling, rate-limit planning, and LLM-as-a-judge evaluations.
Stack
The work demonstrates architecture and implementation depth for AI-platform roles. It does not claim production traffic, SLAs, or model training.