Agentic AI and LLM Applications Software Development Engineer, Senior

Booz Allen Hamilton•Washington, DC

12d•$86,800 - $198,000•Remote

About The Position

The GRACE team at ARPA-H is building the next generation of agentic AI to transform how the agency accelerates research, makes decisions, and ships products at scale. GRACE is ARPA-H's production AI assistant, and we are evolving it into an ecosystem of autonomous, multi-agent systems. We are a small, startup-minded team that ships fast and owns what we build end-to-end. We are looking for a senior SDE who lives at the application layer: designing and building the agentic workflows, LLM integrations, tool-calling systems, and AI-powered features that GRACE users interact with every day. Your focus is on what runs on top of the platform: the agents, the orchestration, the prompts, the pipelines, and the product. The best person for this role starts with the user. They ask why before they ask how. They communicate clearly, give and receive feedback well, and make the people around them better. They are a self-starter with a high bar, a high sense of urgency, and genuine empathy for the people whose work they are making better.

Requirements

7+ years of experience with software engineering, including building and operating production systems
Experience in high-velocity environments where you owned and shipped complex products end-to-end
Experience in Python and at least one other backend language
Experience building and operating systems on major cloud platforms, including AWS, GCP, or Azure
Experience with containerization and working within CI/CD pipelines
Knowledge of modern backend frameworks, async patterns, distributed systems, APIs, data pipelines, and software design patterns
Ability to be a clear, direct communicator who gives and receives feedback well, works with empathy, and makes the people around them better
Ability to be a self-starter with a high bar and high sense of urgency, including not waiting to be told what to do next
Bachelor's degree in Computer Science or Software Engineering

Nice To Haves

Experience building production systems on top of LLMs, including tool-calling, RAG, multi-step reasoning, and context management
Experience with multi-agent (A2A) architectures and orchestration frameworks in production, not just in prototypes
Experience building LLM evaluation and regression testing pipelines
Experience in startup or early-stage environments, including 0-to-1 product building
Experience in big tech building customer-facing AI platforms or developer tools at scale
Experience in security-conscious engineering, including input validation, output sanitization, audit logging, and responsible AI guardrails
Experience in healthcare, life sciences, or other regulated domains
Knowledge MCP at the client/consumer layer, including how agents discover and invoke tools via MCP
Knowledge of token economics, including cost-per-query awareness, context budget management, and prompt efficiency
Ability to demonstrate a strong intuition for prompt engineering and LLM behavior across model families, including why Claude and GPT respond differently to the same prompt and designing for it, and demonstrate comfort with ambiguity

Responsibilities

Design and build GRACE's core agentic workflows: multi-step reasoning, planning, memory, and tool-use across single and multi-agent systems
Implement and evolve A2A communication patterns at the application layer, enabling GRACE agents to collaborate and hand off tasks
Build and maintain the tool-calling layer: tool definitions, input/output schemas, error handling, retry logic, and result formatting
Own the MCP client-side integration: how GRACE agents discover, invoke, and compose tools exposed via MCP servers
Design multi-agent workflows that are reliable, observable, and debuggable in production, not just in demos
Own LLM orchestration at the application layer: prompt construction, context management, model selection logic, and response parsing
Build and maintain RAG features: query formulation, result ranking, citation grounding, and hallucination mitigation
Implement and iterate on prompt engineering patterns and system prompts that drive GRACE's quality and consistency across OpenAI GPT, Anthropic Claude, and Google Gemini
Manage context window budgets: know when to truncate, summarize, or paginate, and build the logic that makes those decisions correctly
Build evaluation pipelines for LLM quality: grounding assessment, regression testing, safety checks, and A/B experimentation on prompt and model changes
Stay sharp on token economics: write prompts and pipelines that are cost-efficient without sacrificing output quality
Translate ambiguous product requirements into clear technical designs and ship them fast
Build new GRACE capabilities end-to-end: from backend application logic through to the API contract the frontend consumes
Rapidly prototype new agentic features, run experiments, collect data, and iterate based on real user behavior
Collaborate closely with product, UX, applied science, and operations; listen well, ask good questions, and build the right thing rather than the obvious thing
Own the quality of what you ship: write tests, handle edge cases, and make sure your features degrade gracefully when upstream dependencies fail
Instrument agentic workflows with tracing, logging, and metrics so failures are diagnosable and regressions are caught before users report them
Define and monitor application-level SLOs: tool call success rates, response quality, and latency from the user's perspective
Build fallback and guardrail logic for AI services: what happens when a model returns something unsafe, off-topic, or structurally wrong
Work closely with the infra engineer to understand system-level constraints and design application behavior that respects them
Write production-quality code: readable, tested, reviewed, and documented
Communicate technical decisions clearly to both engineers and non-engineers; no one should have to guess what you decided or why
Participate actively in design reviews; push back when something is over-engineered or under-specified
Mentor and unblock other engineers; bias toward ownership and fast iteration
Ensure strong privacy, security, and compliance in all application logic and data handling