Derek Wang

Contributed to the core execution paths of the CALL-E goal-oriented outbound voice Agent runtime (120+ Python modules, 35 versioned specs, 33 database migrations), serving 10k+ online users.

Stack: Python, FastAPI, SSE, OpenAI Agents SDK, PostgreSQL (SQLAlchemy AsyncIO / asyncpg / Alembic), Redis, RabbitMQ, Taskiq, Pydantic

• Multi-role Agent boundaries and side-effect gates: MainAgent owns the user chat session (serialized to a single Foreground Turn), while GoalAgent owns long-task lifecycles through an independent goal-scoped session; the two communicate only through Runtime typed events — no SDK handoff or agent-as-tool calls. 100% of real side effects such as outbound calls pass a structured Goal Confirmation gate (a spec-level invariant), with the authorization subject held and re-validated server-side, preventing model-forged authorization.
• Goal → RunSpec → Run durable execution model: Goals own immutable, monotonically versioned RunSpecs, and each Run is pinned to an exact version and input snapshot. The terminal Goal Status and Goal Result (summary + evidence_refs) commit atomically in one PostgreSQL transaction (the Durable Boundary); SSE live delivery over Redis Streams happens only after commit, so lost or duplicated delivery never affects correctness for 10k+ users.
• Execution Lease fencing and Provider Reconciliation: Taskiq workers claim work with leases carrying monotonically increasing fencing epochs, verified exactly on every durable write to fence off zombie workers. Uncertain call submissions are resolved through Provider Reconciliation against a stable provider task name instead of blind redial, driving duplicate calls on the worker-failure recovery path to zero (guaranteed by design).

Core backend contributor for an internal operations Agent platform, reducing end-to-end alert diagnosis from 15–30 minutes to under 60 seconds.

Stack: Java, Spring Boot, Spring AI, DashScope, Milvus, SSE, ReAct, Function Calling

• Three-tier Agent collaboration: designed a Planner → Executor → Supervisor flow around Plan-Execute-Replan. Supervisor validation triggers automatic retries and strategy changes, increasing tool-call success from 70% to 95% across 200+ production alert samples.
• Agentic RAG retrieval: built an internal knowledge base on Milvus with MQE multi-path recall, child-chunk retrieval with parent-chunk returns, Reflection validation, and query rewriting. On 150 internal Q&A evaluations, recall improved by about 25% and retrieval accuracy rose from 65% to 90%.