Multi-agent workflow automation for a two-sided marketplace.
A scaling two-sided marketplace had support costs growing faster than revenue. Buyers waited on shipment and product questions; sellers chased policy and payout answers. Volume of L1 tickets blocked the operations team from scaling without breaking SLAs.
Skip the rip-and-replace. Engineer a permission-aware agent architecture that handles multi-step buyer and seller workflows on top of the existing legacy database, with full auditability and zero downtime cutover.
Enterprise architecture
Distributed multi-agent clusters
Two isolated agent clusters on Kubernetes (EKS).
Buyer agent handles dynamic queries and executes in-chat ordering through generated REST/GraphQL calls.
Seller agent is a permission-scoped RAG copilot reading from internal vector stores (Milvus).
Custom API wrapper middleware
Resilient Go and Node microservices act as a secure bridge: extracted and synced data between the legacy CRM and the new orchestration layer with zero downtime and explicit fault tolerance.
Event-driven orchestration
Kafka manages tool-calls between agents. Strict structured-output validation (Pydantic schemas) gates every action — order placements, refunds, status changes — so downstream effects are deterministic.
DPDPA compliance & localization
All LLM inference and vector search runs entirely within AWS Mumbai. Zero data leaves to public LLM endpoints.
PII redaction pipeline
A pre-processing layer sanitises card numbers, addresses, and phone numbers before any text enters the LLM context window.
Output guardrails
A semantic moderation layer flags hallucinations and policy violations before responses reach the user — no unauthorised commitments.
Human-in-the-loop
High-value refunds and edge cases pause the autonomous flow and escalate to a supervisor for cryptographic approval.
Business outcomes
Fourteen months post-launch, the multi-agent layer runs at 99.99% uptime. As the user base tripled, EKS scaled the clusters automatically without adding a single L1 agent. The custom API wrappers have survived two major legacy-ERP version upgrades without breaking. The in-chat ordering agent now processes millions in monthly GMV — the support cost-centre has become a self-sustaining revenue line.