As an engineer on a small team with end-to-end product ownership, you'll own retrieval, tool orchestration, response generation, and production reliability — from the model-routing layer down to the infra. Design and build the response generation pipeline — routers, orchestrators, synthesizers that classify intent, assemble context, select tools, and decide whether to send or escalate.
Numbers & Facts
Location
San Francisco, California
Description
The build: Our client is building full-stack AI systems for hospitality operations and guest communication. As an engineer on a small team with end-to-end product ownership, you'll own retrieval, tool orchestration, response generation, and production reliability — from the model-routing layer down to the infra.
This is an agent-reliability role: real conversations, messy data, third-party APIs that fail mid-flow. You make it work anyway.
What you'll do
Design and build the response generation pipeline — routers, orchestrators, synthesizers that classify intent, assemble context, select tools, and decide whether to send or escalate.
Develop knowledge retrieval that grounds responses in accurate, complete, current data despite messy inputs.
Build tool orchestration layers that execute multi-step workflows reliably, with graceful handling of API failures mid-conversation.
Create evaluation frameworks that measure quality against real conversations and catch regressions pre-customer.
Implement observability and tracing that makes agent reasoning legible to non-technical support teams.
Design no-code controls that let support leads adjust agent behavior.
Extend the system to voice — real-time transcription and the same tool-use capabilities as text.
Optimize speed and cost via model routing, caching, and prompt refinement.
Own infra, data pipelines, integrations, and API work.
Requirements
What we're looking for
Proven track record shipping production systems real users depend on.
Full-stack comfort — you pick up whatever the problem requires.
Deep thinking about reliable retrieval pipelines.
Pragmatism about models: frontier, fine-tuned, or simpler — whatever fits.
Strong eval discipline; you measure impact on real conversations before shipping.
Experience where your work directly shapes the product