Designing a workflow engine for live agents
Workflow execution semantics
Owned most of the design and implementation of a greenfield workflow engine for live conversational agents. Defined persistent state, validation, retries, approvals, and step-gating semantics. Establishing explicit rules for how an agent could advance or revisit a workflow as the conversation evolved.
Concurrent inference and state updates
Separated response generation and structured extraction into parallel inference paths, keeping extraction off the voice-response critical path. Coordinated tool, CRM, and knowledge-retrieval results before applying workflow-state updates. The design made the ordering of context processing and state updates explicit, accepting additional model cost to avoid serializing response generation behind extraction.
Agent memory and runtime integration
Redesigned cross-call agent memory around a relational fact store with model-driven additions, edits, and deletions after each call. Replaced vector retrieval with explicit fact management aligned to product requirements. Contributed across the shared agent runtime, PostgreSQL/Redis persistence, Celery background workloads, and OpenTelemetry/Grafana observability.