The role
Wilkes & Liberty needs an AI / Agentic Systems Engineer for work where AI is part of the delivered system, not merely an engineering aid. You will own the path from a bounded use case to production behavior: model and retrieval choices, tool authority, evaluation, guardrails, observability, cost, and operator control.
What you will own
- Define the user decision, system boundary, success criteria, unacceptable outcomes, and operator-approval points before selecting a model or orchestration pattern.
- Design and implement retrieval, structured generation, tool use, workflows, and Model Context Protocol integrations with least-privilege authority and explicit data boundaries.
- Evaluate model, provider, and deployment options against quality, privacy, sovereignty, latency, availability, and inference-cost requirements.
- Build representative evaluation sets and automated checks for accuracy, groundedness, refusal behavior, prompt injection, unsafe tool calls, data leakage, and regression.
- Engineer guardrails, schema validation, authorization, rate limits, budgets, audit records, and fail-closed behavior around probabilistic components.
- Instrument model, retrieval, tool, cost, and operator-review behavior so failures can be diagnosed and acceptance claims can be reproduced.
- Plan fallback, rollback, incident response, and change control for model, prompt, retrieval, and tool updates.
- Transfer the system with versioned configuration, tests, evaluation evidence, runbooks, known limits, and clear operator controls.
What you bring
- You have built or operated an AI-enabled product or workflow where model output affected a real user, decision, or system action.
- You can design retrieval and tool-using systems in Python, TypeScript, or a comparable stack and reason about the authority granted at each boundary.
- You treat evaluation as engineering: you define datasets, criteria, thresholds, failure classes, and regression evidence rather than relying on a compelling demonstration.
- You understand model limitations, prompt injection, untrusted content, privacy, access control, observability, and operator oversight.
- You can compare hosted and self-managed models without assuming that the newest or largest model is the correct operational choice.
- You explain quality, cost, latency, risk, and operator-control tradeoffs plainly to technical and non-technical stakeholders.
Helpful experience
- Model Context Protocol, tool calling, structured outputs, retrieval-augmented generation, search and vector retrieval, and evaluation frameworks.
- Open-source or self-hosted models, private infrastructure, air-gapped constraints, and provider-independent system design.
- Identity-aware agent actions, operator-gated workflows, audit logging, and governed access to content or operational systems.
- Federal, regulated, or high-assurance environments where data residency, documented evaluation, and bounded automation determine whether the system is usable.
Equal opportunity and accommodations. Wilkes & Liberty considers qualified candidates without regard to protected characteristics. Reasonable accommodations are available during the application process.