AI Scientist — Agentic Decision Systems

Applied AI Scientist — Agentic Decision Systems

Experience: Typically 2+ years of relevant applied work, although we care substantially more about the quality and difficulty of the problems you have solved than the number of years you have been solving them.

About the Role

We are looking for exceptional quantitative thinkers to build agentic systems that help people make better decisions.

This is not primarily a conventional data science role, an LLM application-development role, or a prompt-engineering role. We are interested in people who want to understand how intelligent systems can reason over complex problems, use tools and models effectively, learn from evidence, and become systematically better over time.

A central question in this role is:

How do we make an agentic system measurably better at the decisions it is intended to support?

That means moving beyond intuition, demos and prompt alchemy. You will think about objectives, observability, feedback, evaluation datasets, statistical uncertainty, experiments, simulation, failure modes and improvement loops. You should be interested not only in building agents, but in developing rigorous ways for those agents—and the systems around them—to improve.

You will work on ambiguous, consequential enterprise problems where the answer is rarely “use an LLM.” Sometimes the right solution will involve an agent. Sometimes it will involve statistical modelling, optimization, machine learning, simulation, rules, or a surprisingly simple piece of mathematics. Frequently it will involve several of these working together.

We want people who are technically ambitious, intellectually demanding, empirically minded and impatient with hand-waving.

What You Will Do

• Design and build agentic decision-support systems that can reason over data, interact with tools and models, maintain state, decompose complex problems and support consequential human decisions.
• Develop mechanisms for systematic improvement. Design evaluation frameworks, feedback loops, benchmarks, experiments and instrumentation that allow us to determine where a system fails, why it fails and whether a proposed change genuinely improves it.
• Turn “make the agent better” into a quantitative problem. Define measurable objectives and failure modes; build representative test sets; analyze performance distributions rather than anecdotes; and distinguish real improvement from movement on a convenient metric.
• Explore self-improving agentic architectures. This may involve reflection and critique, adaptive tool or model selection, memory, search, planning, generated training/evaluation data, policy improvement, human feedback, simulation or other mechanisms. We are interested in what works—not in allegiance to a particular architecture.
• Bring statistical and machine-learning judgment to agentic systems. Use probabilistic modelling, statistical inference, ML, optimization or causal reasoning when they make the system more useful, more reliable or easier to improve.
• Generate high-quality software using AI coding systems. We expect you to work AI-natively. The objective is not to demonstrate how quickly you can type Python; it is to produce excellent software.
• Own difficult problems end to end. Work from an initially ambiguous business or decision problem through problem formulation, system design, implementation, evaluation and deployment.
• Challenge the problem definition. A technically sophisticated solution to the wrong problem is still the wrong solution. You should be comfortable questioning assumptions and reframing what is being optimized.
• Work directly with senior technical and business stakeholders. Explain complex ideas precisely without hiding behind jargon, and make uncertainty, assumptions and limitations explicit.

What We Are Looking For

Exceptional quantitative foundations

You have rigorous training in a highly quantitative discipline such as mathematics, physics, statistics, engineering, econometrics, operations research or another discipline involving substantial mathematical and quantitative reasoning.

A master's or PhD is welcome, but neither is a substitute for evidence of unusually strong analytical thinking. Exceptional candidates with a bachelor's degree are equally of interest.

You should be comfortable reasoning mathematically about unfamiliar problems rather than relying solely on methods you have used before.

Experience with agentic systems

You have meaningful hands-on experience designing or developing agentic systems.

We are particularly interested in experience involving some combination of:

• multi-step reasoning and planning;
• structured outputs and stateful workflows;
• memory and context management;
• multi-agent or decomposed-agent architectures;
• agent evaluation;
• automated critique or refinement;
• feedback-driven adaptation;
• decision-support applications; and
• mechanisms intended to improve system performance over repeated iterations.

A chatbot with a system prompt is not, for our purposes, substantial experience building an agentic system. Neither is wrapping an LLM around a vector database and calling the result an agent.

Statistical and machine-learning literacy

You do not need to be a specialist in every branch of machine learning, but you should have enough statistical maturity to reason properly about evidence.

We expect practical familiarity with statistical modelling and machine learning and, more importantly, sound instincts around:

• uncertainty;
• validation;
• sampling and selection effects;
• overfitting;
• experiment design;
• measurement;
• predictive versus causal claims;
• distribution shift; and
• determining whether a system actually works.

You should instinctively ask what the comparison is, how performance is being measured, whether the evaluation resembles the real problem, and how confident we should be in the result.

AI-native software engineering

You are extremely effective at producing software, but we do not equate software engineering ability with manually writing large quantities of code.

We expect you to use state-of-the-art coding agents extensively—potentially for essentially all implementations.

What matters is whether you can make those systems produce excellent code.

You should be capable of:

• designing a coherent software architecture before or while implementation emerges;
• decomposing work effectively for coding agents;
• inspecting and interrogating unfamiliar generated code;
• identifying bad abstractions, subtle bugs and unnecessary complexity;
• designing meaningful tests rather than merely achieving test coverage;
• maintaining reproducible environments;
• understanding APIs, data pipelines and production system boundaries;
• debugging systems whose implementation you did not personally type; and
• leaving behind code that another strong engineer can understand, extend and trust.

You are accountable for every line you ship even if you personally typed none of them.

The Intellectual Style We Value

The strongest candidates for this role tend to have several things in common.

They are deeply curious but difficult to impress. They enjoy abstraction, but they want abstractions to survive contact with data. They are happy to propose an ambitious architecture and equally happy to discard it when a simpler system performs better. They distinguish a compelling demonstration from convincing evidence. They are interested in why something works, where it stops working, and what experiment would distinguish competing explanations. They are willing to use LLMs aggressively without anthropomorphizing them or treating their behavior as magic. And when someone says, “This version feels better,” their instinct is to ask: “Better according to what measurement?”

You Will Probably Not Enjoy This Role If

• Most of your interest is in conventional analytics, dashboards or notebook-based data science.
• Your experience with generative AI is primarily prompting, chatbot development or basic RAG applications.
• You prefer selecting an AI framework and following its abstractions to thinking from first principles about the underlying system.
• You regard agent behaviour as too unpredictable to evaluate quantitatively or you are satisfied with qualitative demonstrations when controlled measurement is possible.
• You believe increasingly elaborate agent architectures are inherently more intelligent.
• You enjoy writing code but are uncomfortable delegating implementation to AI systems.
• Conversely, you use AI-generated code without wanting to understand its architecture, test its assumptions or take responsibility for its quality.
• You need requirements to be fully specified before you can make progress.

Particularly Interesting, But Not Required

Experience with one or more of the following would be especially valuable:

• decision science, causal inference, Bayesian modelling or optimization;
• reinforcement learning, contextual bandits, simulation or adaptive experimentation;
• agent evaluation, observability, benchmark construction or synthetic task generation;
• inference-time search, test-time computation, model routing or ensemble systems;
• human-in-the-loop or production agent architectures; and
• complex enterprise decision environments such as financial services, insurance, supply chain or logistics.

The role is product-oriented: you will shape the capabilities, workflows and evaluation systems that make our agentic products more useful, reliable and improvable.

You will turn difficult enterprise problems, user needs and client lessons into reusable product capabilities. Working with technical, product, domain and client-facing teams, you will identify high-leverage problems, formulate improvement hypotheses, and design experiments that distinguish genuine progress from superficial gains.

We expect you to move quickly from ideas to prototypes and measurable evidence without compromising reasoning, quality or technical integrity. The goal is not to build the most elaborate AI system, but products that make better decisions, whose performance we can understand, and which we know how to improve.