M
Agentic AI Engineer — Core Product
Multiscale Technologies
31 minutes ago
Full-time
On-site
San Francisco, California, United States
Full-time
The Role
MIND, our agentic platform, powers autonomous agents that design experiments, optimize manufacturing steps, and accelerate scientific discovery. This team has no shortage of ideas - what this seat owns is making them survive production. You'll design the agent harness and evaluation infrastructure from first principles, operate it at scale, and still be running it a year from now. Releases here are human-gated by design: you'll build the machinery that makes an agent's work trustworthy enough for a process engineer to sign off on. You'll partner with our security architecture lead on RBAC, audit trails, and network-isolated serving for air-gapped customer environments - that's a dedicated seat, not this one. What You'll Build
Multi-agent orchestration on LangGraph - sequential, loop, parallel, and supervisor patterns chosen deliberately, deterministic and observable, built to survive real production conditions The operational backbone of long-running agents: workers, queues, timeouts, retries, checkpointing, backpressure, and failure domains Evaluation infrastructure as a first-class system - LLM-as-judge gating, golden datasets, replay runs, and full-stack tracing Human-in-the-loop machinery: checkpoints, review tooling, and judge-gated flows that decide what reaches a person MCP servers and tool-registration APIs that give agents persistent context and shared state across complex workflows Knowledge-graph-backed agent context (Neo4j) and RAG pipelines tuned for materials-science and fab retrieval The API contracts with the core Multiscale platform - the foundation everything else depends on What We're Looking For
You've designed a production system from a blank page - chose the queues, workers, timeouts, retries, and failure domains yourself - and can defend every trade-off You've operated an agentic or LLM system at scale - on call for it, with real stories about resilience: checkpointing, idempotency, backpressure, recovery You know what's under the hood - transformer mechanics, KV cache, context and inference costs - well enough to reason about latency, memory, and failure, not just call APIs You can compare agent orchestration patterns - sequential, loop-based, parallel, supervisor - and argue when each is the wrong choice, including when not to use an agent at all You've gated something real on evaluation - judge thresholds, golden-set regressions, replay runs - and evals show up in your answers before we ask 5+ years of production software engineering, including 2+ years shipping LLM or agentic systems; Docker and Kubernetes are second nature on Azure, AWS, or GCP What Will Make You Stand Out
Programmatic prompt optimization (DSPy or similar) Uncertainty quantification or model calibration work Graph databases in production Network-isolated or regulated-environment deployment experience Any exposure to semiconductor manufacturing, fab operations, or materials science Open-source agentic work - show us the system you operated longest, not the flashiest demo Our Process
Intro call, then a technical round: rapid fundamentals plus a from-scratch systems design - expect us to change the requirements mid-design and to care about the questions you ask before you build. Decision follows quickly. Why Join Us
Own subsystems end to end and run them release after release - real ownership, not handoffs About 30 people, fast decisions, many hats, and direct access to customers and their hardest problems Work at the intersection of frontier AI and one of the world's most complex engineering domains Competitive compensation with meaningful equity
#J-18808-Ljbffr
MIND, our agentic platform, powers autonomous agents that design experiments, optimize manufacturing steps, and accelerate scientific discovery. This team has no shortage of ideas - what this seat owns is making them survive production. You'll design the agent harness and evaluation infrastructure from first principles, operate it at scale, and still be running it a year from now. Releases here are human-gated by design: you'll build the machinery that makes an agent's work trustworthy enough for a process engineer to sign off on. You'll partner with our security architecture lead on RBAC, audit trails, and network-isolated serving for air-gapped customer environments - that's a dedicated seat, not this one. What You'll Build
Multi-agent orchestration on LangGraph - sequential, loop, parallel, and supervisor patterns chosen deliberately, deterministic and observable, built to survive real production conditions The operational backbone of long-running agents: workers, queues, timeouts, retries, checkpointing, backpressure, and failure domains Evaluation infrastructure as a first-class system - LLM-as-judge gating, golden datasets, replay runs, and full-stack tracing Human-in-the-loop machinery: checkpoints, review tooling, and judge-gated flows that decide what reaches a person MCP servers and tool-registration APIs that give agents persistent context and shared state across complex workflows Knowledge-graph-backed agent context (Neo4j) and RAG pipelines tuned for materials-science and fab retrieval The API contracts with the core Multiscale platform - the foundation everything else depends on What We're Looking For
You've designed a production system from a blank page - chose the queues, workers, timeouts, retries, and failure domains yourself - and can defend every trade-off You've operated an agentic or LLM system at scale - on call for it, with real stories about resilience: checkpointing, idempotency, backpressure, recovery You know what's under the hood - transformer mechanics, KV cache, context and inference costs - well enough to reason about latency, memory, and failure, not just call APIs You can compare agent orchestration patterns - sequential, loop-based, parallel, supervisor - and argue when each is the wrong choice, including when not to use an agent at all You've gated something real on evaluation - judge thresholds, golden-set regressions, replay runs - and evals show up in your answers before we ask 5+ years of production software engineering, including 2+ years shipping LLM or agentic systems; Docker and Kubernetes are second nature on Azure, AWS, or GCP What Will Make You Stand Out
Programmatic prompt optimization (DSPy or similar) Uncertainty quantification or model calibration work Graph databases in production Network-isolated or regulated-environment deployment experience Any exposure to semiconductor manufacturing, fab operations, or materials science Open-source agentic work - show us the system you operated longest, not the flashiest demo Our Process
Intro call, then a technical round: rapid fundamentals plus a from-scratch systems design - expect us to change the requirements mid-design and to care about the questions you ask before you build. Decision follows quickly. Why Join Us
Own subsystems end to end and run them release after release - real ownership, not handoffs About 30 people, fast decisions, many hats, and direct access to customers and their hardest problems Work at the intersection of frontier AI and one of the world's most complex engineering domains Competitive compensation with meaningful equity
#J-18808-Ljbffr