Member of Technical Staff, Core Backend

Vapi•San Francisco, CA

1d•$200,000 - $280,000

About The Position

Voice AI that resolves, not transfers. Most phone systems trap callers in menus and scripts. Vapi is the platform for deploying voice agents that know your business and can listen, adapt, and resolve in minutes. The numbers: 1 billion calls. 1 million developers. 10x enterprise ARR growth The customers: Amazon Ring, ServiceTitan, New York Life, Intuit, Kavak, and thousands more, from YC startups to the Fortune 500 The news: a $50M Series B led by Peak XV Partners, with Bessemer Venture Partners, Kleiner Perkins, M12 (Microsoft's Venture Fund), Y Combinator, and our earlier backers. Total raised: $72M Why We’re Hiring This Role: The StreamModule pipeline — VAD → STT → LLM → TTS → Transport — runs on cork/uncork backpressure during live phone calls. A hundred milliseconds of delay is audible. This role owns pipeline stability and pluggability, so the agents and FDE teams can add new models and providers without touching core. You’ll consolidate BullMQ into Kafka, harden the provider abstractions (LLM, STT, TTS base classes), instrument the pipeline with event-driven OTEL tracing, and shore up the Postgres SPOFs that contributed to the Oct 15 and Oct 22 incidents.

Requirements

Built real-time or streaming systems in production — media pipelines, streaming data, or event-driven backends.
Debugged a backpressure cascade.
Opinions on queue architecture (BullMQ, Kafka, Temporal) and when each is the right fit.
Built plugin or adapter architectures — extending base classes cleanly, with decoupled implementations.
Operated Postgres at scale: connection pooling, read replicas, schema migrations (Liquibase or similar).
Instrument with OpenTelemetry and think in event-driven traces, not just logs.

Nice To Haves

TypeScript + Node.js + NestJS.

Responsibilities

Ramp on the StreamModule pipeline and the cork/uncork backpressure model.
Walk the Oct 15 / Oct 22 DB incidents and the duplicate-message incident.
Land a scoped pipeline or provider-abstraction improvement.
Own a slice of the BullMQ → Kafka consolidation.
Ship event-driven OTEL instrumentation for at least one critical pipeline stage.
Harden one provider plugin path so a new model can be added without core changes.
Drive a measurable reliability or latency win on the call path.
Be the backend owner that agents and FDE teams pull in for design reviews on new providers and pipeline changes.