Frontier AI companies are increasingly bottlenecked on expert judgment and high-quality data workflows. This team builds the production systems that capture, coordinate, and validate that work at scale — directly between a customer request and the output that ships
These are long-running, stateful systems. A single job can stay live for days, interleaving automated steps, model inference, and expert review. A step marked "done" can be reopened, re-reviewed, and redone — so "completed" is not always final, state has to tolerate late mutation, and correctness has to survive humans and models disagreeing with each other
This is a backend systems and orchestration problem: distributed state machines, not pipelines. The architecture is not set. Early engineers will decide what it becomes, and the loop between "I shipped this" and "this mattered" is short
Design services and state models for multi-stage workflows that fan out across automated processing and expert reviewers, then reconcile results into a coherent whole
Build orchestration primitives — retries, failure recovery, idempotency, auditable state transitions — for jobs that run far longer than a request and can be partially redone after the fact
Integrate model inference into production workflows without sacrificing debuggability or human oversight
Build the APIs and tooling that let product, operations, and ML teams operate, debug, and trust these systems at scale
Own reliability and observability for workflows where a silent failure means a corrupted result, not just a 500
Moving fast on genuinely hard systems problems — ambiguity is the default, not the exception
Working closely with product, operations, and ML teams to translate a tangle of constraints into clean system design
Debugging complex stateful workflows where the failure surface spans automated steps, model calls, and human reviewers
Owning your systems end-to-end: design, ship, operate, improve