How Applied Science Leaders Build Teams That Compound

After building applied science teams across Amazon, Walmart, Adobe, Salesforce, and Microsoft — across search, recommendations, document AI, marketing measurement, and generative AI — I have found that the role is best defined the way Amazon defines it: the combination of deep scientific expertise, production engineering capability, and the ability to decompose complex business problems into components that science can actually solve.

The Applied Science Leader owns all three simultaneously. They are not pure researchers generating knowledge for its own sake. They are not pure engineers implementing known solutions. They are the function that translates between what is scientifically possible and what actually ships.

What Amazon’s definition does not tell you is how to build the team that makes it real. This essay is that guide.

Start With the Diagnostic

The most common mistake an Applied Science Leader makes in the first thirty days is starting to hire immediately. Headcount feels like action. It is not. Hiring before you understand what you are building for is the fastest way to assemble a team that is well-credentialed, well-intentioned, and structurally wrong for the work.

McKinsey’s AI research categorizes organizations into three levels of maturity: Starters, Experimenters, and Leaders. The Applied Science Leader’s first job is to determine honestly which world they have walked into — because the team shape, the hiring sequence, and the first year’s priorities are fundamentally different in each. A Starter needs foundations before it needs scientists. An Experimenter needs architectural correctness before it needs scale. A Leader needs an honest audit of what is genuinely compounding versus what is demoware accumulating as technical debt.

That determination requires a structured assessment. What has the organization actually shipped — not what is on the roadmap but what is in production, how long it has been running, and whether it was built correctly or cobbled together for a demo? Who owns the data pipeline — is the science team a first-class stakeholder or a consumer of another team’s decisions? What compute and ML infrastructure exists? What does the evaluation framework look like — rigorous ground truth and failure mode coverage, or a dashboard of engagement metrics a PM reviews? Can the team observe model behavior in production or does degradation happen silently? What do customer experience reports reveal about where the current AI capability actually breaks?

None of these questions have comfortable answers in most organizations. That discomfort is information. The diagnostic is not complete until the Applied Science Leader can answer all of them honestly — and has mapped the gap between where the organization is and where the team needs to take it.

Build the Composition Deliberately

Once the diagnostic is complete, team composition becomes a principled decision rather than a guess.

A team doing novel modeling on ambiguous problems needs higher research density — more principals and senior scientists who can define the problem correctly before attempting to solve it. A team improving and scaling a well-understood system needs higher engineering density. The error is hiring the wrong profile for the phase. A team of researchers handed a scaling problem will produce papers. A team of engineers handed a novel problem will produce confident wrong answers quickly. Knowing which phase you are in — and which phase is coming — is the Applied Science Leader’s job.

In the generative AI era the Evaluation Lead is not a support role. It is arguably the most consequential hire on the team. Without rigorous evaluation infrastructure everything the team ships runs on organizational faith rather than evidence. This person is not a junior annotator. They are a senior scientist with deep experimental design expertise, domain knowledge sufficient to define what failure means in the specific problem space, and the professional independence to surface findings nobody wants to hear. They should be among the highest compensated and most structurally empowered people on the team — the organizational equivalent of internal affairs.

The seniority composition decision should start from problem requirements and then negotiate with budget — not the other way around. A team built primarily from juniors to save money will spend two years without architectural foundations. A team being built from scratch needs higher senior density to establish norms, evaluation standards, and the institutional memory that only comes from having made and learned from the foundational mistakes.

Instrument the Hiring Funnel

Most Applied Science Leaders inherit a broken hiring funnel and accept it as given. Start recording data immediately — resume submissions, filtering rate, manager screen pass rate, onsite conversion, offer rate, accept rate. The target is roughly one hire per three onsites. If you are hiring one in ten, the early screens are not working. If you are hiring two in three, the bar is too low.

The manager screen should have one specific goal: identify candidates with a high probability of passing the onsite. Ask one technically substantive question — not a coding problem, not a brain teaser, but a question that requires the candidate to reason at the intersection of their claimed expertise and the actual work the team does. How they handle that question tells you more than the rest of the conversation combined.

The best applied scientists do not all have the same pedigree. The signal is not the credential. It is the combination of genuine depth in at least one domain, the ability to reason from first principles in unfamiliar territory, model shipping experience, and intellectual hunger. The institutional brand shortcut — filtering heavily on school or company name — systematically excludes the best candidates from institutions that do not carry marquee names. Some of the strongest applied AI talent comes from exactly those places.

Test for Cognitive Traits, Not Credentials

Each level of seniority requires testing for a fundamentally different cognitive trait. The same interview with a higher bar is not a senior interview — it is a junior interview that filters out everyone.

For a principal, mathematical precision and architectural rigor are non-negotiable. Start with a concept they know deeply. Get a fast precise answer. Then push them into uncomfortable territory — ask them to extend the math to a domain they have not worked in, or redesign a core mechanism under different constraints. They do not need to know the answer. They need to demonstrate the thought process. The grandmaster test is whether they can hold the entire stack in their head simultaneously — the relationship between data quality, model architecture, evaluation design, and production behavior as an integrated system. The adaptation test is the most important signal: when you introduce a constraint that invalidates their approach, do they find it interesting or do they defend the original answer? A principal who defends is telling you how they will behave when their architectural decisions are challenged in production.

For a senior, you are testing for direction and intellectual hunger. Do they have a clear sense of what they want to work on and why? Can they articulate the difference between a method that works and a method that is right — and do they care about the distinction? The best signal: they ask you hard questions during the interview.

For a junior, you are hiring for trajectory. Do they have real depth in one area or broad shallow exposure to many frameworks? A junior who truly understands one area deeply and knows how to use modern tools to scale their output is infinitely more valuable than one who has touched everything and understood nothing.

Build the Culture That Attracts the Next Hire

The hiring funnel is not only a pipeline — it is a reputation. The strongest applied science teams fill a significant portion of their roles through referrals from existing team members. Strong scientists know other strong scientists. They refer the people they respect to the environments they respect. The culture the Applied Science Leader builds determines whether those referrals happen.

The culture that attracts top talent has specific properties. It keeps the work at the frontier — there is always a problem worth thinking hard about, always a question that requires genuine exploration rather than execution of a known pattern. It protects the shipping cadence — research feeds shipping and shipping feeds research, and the team can point to real capability in production rather than a permanent backlog of promising experiments. It models intellectual honesty — the Applied Science Leader is the first person to say publicly when an approach they backed was wrong, and the team learns that surfacing a problem is rewarded rather than punished.

The IC track matters enormously. Strong scientists often do not want to manage. If the only visible path to seniority runs through people management, the ones who want to go deep will leave for organizations that have built and funded a principal and distinguished scientist track. Establishing that track — with real compensation attached, not just a title — and visibly celebrating the technical contributions that earn progression on it is one of the most powerful retention and referral mechanisms available.

A culture that supports publication signals to the broader community that this is a place where real work happens. A senior scientist who gets a paper accepted at NeurIPS and stays because the environment supports that work will refer two others like them. The referral network is the best recruiting channel available and it is built entirely from the inside out.

The wrong hire damages this as much as the right hire builds it. Wrong hire signals are almost always visible within sixty to ninety days — the person who pattern-matches rather than reasons, whose outputs look right on the surface but do not hold up under technical scrutiny. Acting on those signals early — through honest feedback, reallocation, or a managed exit — is not harsh. It is the most important thing the Applied Science Leader can do for the team’s culture and for the referral network that feeds future hiring.

The Compounding Team

A well-built applied science team does not produce linear returns. It compounds. Each strong hire raises the bar for the next hire. Each rigorous evaluation framework catches failure modes that would otherwise have shipped. Each principal who sets the intellectual standard elevates the reasoning quality of every scientist and engineer around them.

The compounding effect takes time to become visible. In the first six months the team looks like cost. In the second year it looks like capability the organization cannot explain — models that actually generalize, evaluation pipelines that catch failures before they reach customers, research directions that prove right for problems the business did not know it had yet.

The Applied Science Leader’s job is to build the team that produces the second year’s results — and to protect the first year’s investment long enough for those results to arrive.

The team is the first product. Everything else follows from it.

If you are building or restructuring an applied science team and want to compare notes — I am always happy to connect.