Applied AI Engineer

Nxt Level; New York (NY); On-site; USD 180,000 - 240,000; Full time

Job summary

Nxt Level is seeking a Member of Technical Staff focused on Applied AI to own our AI stack end to end. You’ll design and implement systems where agents are the primary users of our system of record, ensuring reliability across data models, APIs, and the UI.

You’ll build a general agent capable of multi-step tasks, ambient agents that triage work, and an orchestration layer that brings LLMs, context, tools, and business logic into a coherent experience.

Job description

As a Member of Technical Staff focused on Applied AI, you’ll own our AI stack end to end. One framing we keep coming back to: agents are the primary users of our system of record. Everything we build (the data models, the APIs, the UI) has to work for a non-human user that operates at scale, across every client, all the time. That’s a different design constraint than most teams are used to.

What you’ll build (and own)

• A general agent capable of complex, multi-step tasks — planning, sandboxed code execution, web search, retrieval — that powers a "do anything" experience for advisors.
• Ambient agents that act on behalf of clients and advisors: triaging email, processing meetings, drafting communications, surfacing what needs attention before anyone asks.
• The agent harness that orchestrates LLMs, context, tools, retrieval, and business logic into something coherent and reliable.
• Generative UI and human-in-the-loop interfaces where the agent and advisor genuinely collaborate, not just take turns.
• Evaluation infrastructure that holds two bars at once: high-correctness financial data and subjective, judgment-heavy tasks.

Example problems you’d work on

These aren’t hypothetical problems; we’re actively working on versions of all of these.

• Human / AI collaboration that actually works in practice. An advisor is mid-call when the agent surfaces a multi-step recommendation: rebalance, adjust the savings rate, revisit the estate plan. The advisor takes two steps and overrides the third. Now what? How does the agent update its model of what this advisor wants, present reasoning the advisor can relay without sounding scripted, and learn over time what to do autonomously versus flag? This is the flywheel: the tighter the collaboration, the more the agent can take on.
• Memory systems that know a client the way a great advisor does. A good advisor remembers that a client gets anxious when markets drop, cares more about the kids' college fund than their own retirement, and prefers plain-English summaries. Building memory that captures and evolves this understanding across years—and surfaces the right context at the right moment is genuinely hard. The challenge is knowing what to retrieve, what's still relevant, and how to represent a person's relationship to money in a way an agent can use.
• Generative UI as an agent architecture problem. When an agent views and updates the advisor's screen in real time—rendering scenarios, adjusting visualizations mid-conversation, surfacing recommendations inline—the UI is constantly changing. The challenge is how the agent manages state across those changes and how you keep the experience from feeling unpredictable. When the visualization shows a client's actual retirement savings, the advisor can't be surprised by their own screen.
• Evaluations that work for financial services. Most evals are built for tasks with a single right answer. Financial advice isn’t like that; the same recommendation can be right for one client and wrong for another, and “correct” often depends on context the eval harness doesn’t have. You’ll build evaluation infrastructure that can hold two different bars simultaneously: high-accuracy financial data where errors have real consequences, and judgment-heavy tasks where the right answer is subjective and the stakes are relational. Add to this the compliance requirements of financial services, where auditability isn’t optional and infrastructure has to handle large, constantly changing datasets where a stale answer can be as harmful as a wrong one.

The kind of person who thrives here

You’re excited by ownership, ambiguity, and building things that matter.

• You're comfortable where "correct" isn't always obvious. Financial advice isn't deterministic, and neither is evaluating it. You're energized by probabilistic systems and rigorous about the evaluation infrastructure that tells you whether you're improving—you trust experimentation more than your first instinct.
• You're ambitious in a way specific to this work. You're building agents that touch real retirement savings and estate plans. A hallucination here isn't a product bug; it's a wrong answer that affects someone's financial future. That consequence makes you more careful, not slower.
• You move fast, and you know speed and reliability aren't in tension here. An agent that behaves differently in production than in eval is a liability. You treat evaluation and iteration as part of shipping, not steps that come after.
• You have taste, and a high bar for what that means here. You spot AI slop instantly and won't let it through review. You know "tech debt vs. shipping" is a false tradeoff and act accordingly—leaving systems better instrumented and better understood than you found them.
• You’re someone people actually want to be in the room with. This is a 12-person team, one office, five days a week, working on problems that don’t have clean answers. The people who thrive here to argue about agent architecture at lunch and then run an eval together in the afternoon. Kind, direct, and low-ego, you can give candid feedback without being an asshole, and you’re genuinely energized by this environment (not just tolerant of it).