Staff Machine Learning Engineer, ServiceNow

• 8+ years building production software, with several years specifically shipping LLM-powered / agentic systems (not just API wrapper calls — real tool-use loops, state management, multi-turn orchestration)
• Experience in leveraging or critically thinking about how to integrate AI into work processes, decision-making, or problem-solving. This may include using AI-powered tools, automating workflows, analyzing AI-driven insights, or exploring AI's potential impact on the function or industry
• Deep, hands-on expertise with LangGraph and/or LangChain (or the judgment to know when to skip them and hand-roll something better)
• Strong understanding of MCP — ideally having built an MCP server, not just consumed one
• Production async Python (FastAPI, asyncio) — comfortable with WebSockets, streaming, and the failure modes of long-lived stateful connections
• Track record of making real architecture decisions on agent systems — tool boundaries, context/state design, cost/latency tradeoffs, when a bounded agent beats a fully autonomous one
• Enough range outside Python to read/modify a Spring Boot service and a React component without hand-holding — this is explicitly a whole-stack role, not "Python specialist with a frontend allergy"
• Comfort operating with ambiguity and setting technical direction, not just executing a spec — this is a Staff-level bar on judgment
• A demonstrated habit of pulling the latest from the industry — new agent frameworks, protocol standards, model capabilities — into production quickly
• Willingness and ability to build real fluency in the Core ServiceNow platform. Our systems increasingly need to interoperate with it directly, and this role is expected to help drive that, not just react to it
• Willingness to work directly with customers/deployments as part of Forward Deployed Engineering efforts — this isn't a purely internal-platform role

Desirable

• We're looking for someone who already operates at a Senior-Staff bar in the agentic/LLM domain but is building out breadth across the rest of the stack
• You'll be one of the most senior technical voices on how agentic systems get designed here — state management, tool boundaries, streaming protocols, prompt/context architecture, and multi-agent coordination — while staying credible end-to-end: able to read a Spring Boot service, unblock a frontend integration, or reason about a classical ML model pipeline when the problem calls for it
• Multi-agent orchestration — LangGraph/LangChain agents over frontier LLMs for transaction editing, conversational configuration, and multi-product quote planning with plan/approve/refine loops and parallel task execution
• The Harness — we're crystallizing our own agent execution runtime into an industry-leading, state-of-the-art harness. Full-duplex sessions where a user can interrupt, redirect, or answer a clarifying question mid-execution while other work keeps streaming, built on a from-scratch async runtime rather than a bolted-on wrapper around someone else's agent loop. This is as much a performance and UX problem as a backend one — low-latency streaming, backpressure, live progress, partial results, graceful cancellation — and it's the part of the stack we're most invested in owning outright. You'd be a primary owner of where this goes next
• MCP as a secondary interface — we maintain a first-party MCP server and clients into our admin/product/rules/transaction systems, but as the Harness matures it becomes the primary way our own agents interact with the platform, with MCP kept as the secondary, standards-based surface for external interop. You'd help decide what stays MCP-first and what moves onto the Harness
• A2A protocol — agent-to-agent task delegation and streaming, surfaced through an external gateway so other systems (including core ServiceNow) can drive our agents directly
• Forward Deployed Engineering — expect real time embedded with customer- and product-facing teams against live deployments. Adapting the Harness and our agents to actual customer workflows under real constraints, not just building platform capability in the abstract
• RAG / context engineering — tenant-uploaded document ingestion, categorization, and aggregation into agent context. Prefix-cacheable prompt design for cost/latency
• Classical ML, when the problem isn't a good fit for an LLM — we have a separate PyTorch/scikit-learn training and serving pipeline (field-value prediction) that a whole-stack ML engineer should be able to read, extend, or evaluate against LLM-based alternatives
• Full-stack fluency — enough comfort in Spring Boot/Java services and the React/TypeScript + Lit frontend to unblock an integration end-to-end without waiting on a handoff