Sr Staff Software Engineer

Role summary The Senior Staff Software Engineer (IC5) on HR Service Delivery sets and owns the technical direction for model-driven capability across the HR domain and the surfaces adjacent domains build on top of it. The work is agentic and conversational experiences that interpret an employee's, manager's, or HR agent's intent, reason over profile, case, catalog, policy, and knowledge context, invoke tools, and act on the user's behalf — but the unit of ownership is the architecture, the standards, and the evaluation infrastructure that let multiple teams ship those experiences safely, not a set of features. This is not a machine learning or AI research role; the engineer does not train foundation models. It is also distinct from traditional staff-level full-stack engineering, where systems follow deterministic logic rather than selecting execution paths at runtime. Two consequences shape the work. First, the most important logic often lives in natural language — instructions, prompts, tool descriptions, guardrails, escalation rules — which must be engineered, versioned, and reviewed with the same discipline as code, and which needs an architecture once more than one team is authoring it. Second, because behavior is probabilistic, correctness is established by measuring behavior at scale rather than by asserting fixed outputs, which makes evaluation a platform investment rather than a per-feature activity. HR sharpens both points along two independent axes. Accuracy. An agent answering on payroll, benefits, leave, or a lifecycle event touches statutory entitlement, jurisdiction-specific policy, and an employee's pay. A confidently wrong output is not a bad answer; it is a missed enrollment window or an incorrect leave balance acted on in good faith. Audience. HR data is among the most sensitive on the platform, and correctness of content is not sufficient. An answer grounded in a record or knowledge article the requestor is not entitled to read is a data exposure even when every fact in it is true, and manager-scope and employee-scope views of the same question have different correct answers. Enforcement belongs in the retrieval and tool layer, not in a request to the model, and at this level you own that being true by construction across every team building in the domain. The distinguishing expectation at IC5 is that the hardest problems arrive unframed. You decide what the domain should do about them, commit the organization to an approach, and are accountable for that approach across releases — including for the decisions that turn out wrong. What you do Set the AI architecture for the domain Own the architecture of model-driven capability across the employee lifecycle: how agents are decomposed and composed, where reasoning happens, how context is assembled and bounded, how tools are exposed and described, and how autonomy is delegated and revoked. Establish the reference patterns other teams build against, and the boundaries between what is a shared platform concern and what each product area owns. Make the calls with multi-release consequences — model selection and migration, orchestration approach, build-versus-adopt, and the cost, latency, and quality tradeoffs behind each — and own the outcome. Define where autonomy goes, and where it does not Decide, as a matter of domain policy rather than per-feature design, which HR actions an agent may take, which require a human decision point, and which no agent should attempt. Anything that changes pay, employment status, or a restricted record, and anything touching employee relations or investigation, needs a human in the path by construction. Then build the mechanisms that make those constraints structural and hard to violate accidentally, so that a team shipping a new capability inherits the boundary instead of re-deriving it. You hold the authority to refuse a shipping decision on these grounds, and are expected to use it. Own the shared instruction and tool-description