About the role
Between a user's question and a verified answer sits an agent that must plan a campaign, pick which businesses to call, hold conversations, negotiate, recover from dead ends, and know when it's done. You'll own that loop.
This is frontier work: long-horizon agents acting in the real world, where a hallucination isn't a bad paragraph — it's a wrong booking. You'll design the guardrails, evals, and memory that make Aurora trustworthy at scale.
What you'll do
- Design the agent architecture — planning, tool use, conversation state, and multi-call orchestration.
- Build evaluation suites that measure task completion, truthfulness, and negotiation quality.
- Own prompting, fine-tuning, and model-routing decisions across providers.
- Turn messy real-world call transcripts into structured, verified answers users can act on.
What we're looking for
- You've shipped LLM-powered systems to production — agents, RAG, or tool-use pipelines.
- Strong engineering fundamentals; you treat prompts and evals as code.
- Comfort with ambiguity: the right architecture doesn't exist yet, and you like that.
- You measure before you believe — evals over vibes.
Nice to have
- Experience with multi-step / long-horizon agent frameworks.
- Background in RL, planning, or dialogue systems.
- Prior founding-team or 0→1 experience.