Advisory Software Engineer

Own and groom the backlog for one or more AI pods (agent platform, AI advisor products, or AI productivity tooling). Set sprint goals that ladder to quarterly outcomes, not activity.

Translate ambiguous executive intent ("we need an agent that helps SMBs ship smarter") into epics, stories, and acceptance criteria engineers can start on Monday.

Write acceptance criteria for probabilistic systems: define what "good" means for an agent response, what the eval set is, what the pass threshold is, and what the failure mode is when it misses.

Own the definition of done for agent capabilities - including evals, guardrails, observability, cost per interaction, and rollback path, not just "the happy path works."

Prioritize ruthlessly across competing stakeholders: product management, engineering leadership, security, architecture, and the business units consuming the platform.

Maintain the tool and capability catalog for the agent platform - which APIs are exposed as agent tools, which are approval-gated, which are read-only, and why.

Run sprint planning, standup, review, and retrospective for the pod. Keep them short and keep them useful.

Track and report delivery health: velocity, cycle time, spillover, blocked-time. Bring problems forward early rather than explaining them in hindsight.

Remove impediments - cross-team dependencies, environment access, security review queues, vendor bottlenecks. Escalate with a proposed resolution, not just a flag.

Facilitate estimation and scope negotiation in a domain where estimates are genuinely uncertain, without letting uncertainty become an excuse.

Coordinate across pods and with partner teams (platform engineering, data, security, product) so integration work is planned rather than discovered.

Partner with the architecture lead on target-state decisions and carry those decisions into the backlog with enough fidelity that they survive contact with implementation.

Maintain requirement traceability from product requirements through to delivered agent behavior, including where a requirement was deliberately descoped and why.

Own the AI governance checkpoints in the delivery flow: model approval, data handling review, human-in-the-loop placement, prompt and tool change control, and audit evidence.

Keep a live view of platform economics - token spend, model selection, inference cost per use case - and treat cost regressions as defects.

Run POCs as time-boxed experiments with a written decision at the end, not as open-ended projects.

You know the platform architecture well enough to explain it to a VP without an engineer in the room.

The backlog for your pod is groomed two sprints deep with acceptance criteria that engineers do not have to re-litigate in planning.

You have identified and closed the three largest sources of delivery friction for the pod.

Predictable delivery: sprint goals met consistently, spillover trending down, dependencies surfaced before they block.

Every shipped agent capability has an eval set, a guardrail spec, and a cost-per-interaction number attached to it.

Stakeholders across product, engineering, and security come to you for status rather than assembling it themselves.

Agent runtime and orchestration - multi-agent systems on Bedrock AgentCore: routing, session and memory management, tool selection, retry and fallback behavior, and human-in-the-loop approval gates for actions that cost money.

MCP tool servers - clean, well-scoped tool interfaces over production APIs (address validation, rate shopping, label creation, tracking). Deciding what not to expose is as much of the job as building what you do.

Retrieval and grounding - ingestion pipelines, chunking and embedding strategy, vector and hybrid search, and freshness guarantees over documentation, contracts, and operational data.

Evaluation and guardrails - golden datasets, automated regression gates in CI, LLM-as-judge harnesses, input/output filtering, PII handling, and prompt-injection defenses on any tool that touches untrusted content.

Platform plumbing - auth and identity propagation across agent and tool layers, rate limiting, caching, cost and token instrumentation, tracing, and the dashboards that tell us when quality has quietly regressed.

5+ years in technical product ownership, technical program management, or engineering delivery leadership on software platforms - with at least 2 years directly accountable for a backlog.

Demonstrated experience delivering AI/ML or LLM-based features to production. Prototypes and pilots count only if you can describe what broke when real users arrived.

Working fluency with modern AI application patterns: prompting, RAG, tool/function calling, agent orchestration, evaluation, and guardrails. You do not need to write the code; you need to reason about the design.

Strong API literacy - you can read an OpenAPI spec, understand auth models, and reason about latency, idempotency, and error handling.

Proven Scrum or Kanban facilitation with distributed teams, including offshore or multi-timezone pods.

Excellent written communication. This role produces a lot of writing that executives read.

Comfort with ambiguity and with saying "not this sprint" to senior stakeholders.

Pitney Bowes (NYSE: PBI) is a global shipping and mailing company that provides technology, logistics, and financial services to more than 90 percent of the Fortune 500. Small business, retail, enterprise, and government clients around the world rely on Pitney Bowes to remove the complexity of sending mail and parcels. For additional information visit Pitney Bowes at www.pitneybowes.com .

• Provide the will: opportunity to grow and develop your career • Offer an inclusive environment that encourages diverse perspectives and ideas • Deliver challenging and unique opportunities to contribute to the success of a transforming organization • Offer comprehensive benefits globally ( PB Benefits and Wellbeing Programs )

Pitney Bowes is an equal opportunity employer that values diversity and inclusiveness in the workplace. All interested individuals must apply online.