Agentic AI
B2B Customer Service Voice Bot 2.0
Confidential Client Builds
Customer service modernization for a B2B apparel and distribution business unit.
LangGraphLangChainVertex AIGKERAGVoice Bot
Architecture Responsibility
Responsible for technology architecture and hands-on delivery direction across system design, deployment, DevOps, cost, scale, reliability, and production readiness.
Outcome
Modernized B2B customer service by automating complex inquiries across inventory, pricing, order tracking, and shipping. The system targeted 80% automation, reduced live agent load, and supported English and Spanish interactions.
Scale
Designed for high availability and handled approximately 1,000 calls per month.
Architecture
- Moved from a rigid Dialogflow-only implementation to a multi-agent architecture using LangGraph and LangChain.
- Hosted the agentic system on a multi-node Google Kubernetes Engine cluster with multi-zone redundancy.
- Used Vertex AI Gemini for generative responses and Vector Search for RAG-based FAQ retrieval.
- Connected securely to legacy on-premise ERP systems through VPN/private tunnels for real-time lookups.
- Implemented a dedicated PII redaction layer before persistence or logging.
Lessons Learned
- Intent-only bot architectures become brittle as business workflows grow; agentic systems need modular boundaries that can be observed and debugged.
- Production LLM systems require traces, metrics, and prompt/output visibility from day one, otherwise reliability and cost cannot be managed.
- For enterprise AI, architecture must define safe limits clearly: read-only actions, escalation paths, and human control for sensitive workflows.