09 · Career Growth: IC to Principal Data Scientist

The skills that get someone promoted from junior to senior (technical correctness, independence) are not the same skills that get someone from senior to staff/principal (scope, leverage, organizational judgment). This module covers what actually changes at each level and how to build evidence of it deliberately, rather than waiting for it to be recognized.

What actually changes across levels

Junior:    Executes well-scoped tasks with guidance. Correctness is the bar.
Mid:       Owns a project end-to-end. Scopes its own smaller tasks.
Senior:    Owns ambiguous problems. Decides WHAT to work on, not just how.
           Technical judgment trusted without close review.
Staff:     Influences direction across multiple teams without direct
           authority over them. Work often looks like: identifying a
           problem no one assigned, building consensus, and having
           OTHER people execute much of it.
Principal: Sets technical/strategic direction at an org level. Judged on
           org-wide outcomes, not on projects with your name on them.

The senior→staff transition is where most data scientists get stuck, and it's usually not a skills gap — it's that staff-level work often produces less individually-attributable output (you convinced three teams to adopt a shared metric definition; you didn't ship a model) which is harder to put in a promotion packet than "I shipped model X."

Building evidence of scope, not just skill

Weak promotion evidence: "I'm very good at causal inference."
                          (a skill claim — the bar for this is already
                          "correct," it doesn't differentiate levels)

Strong promotion evidence: "I noticed three teams were computing 'active
                            user' differently, which was causing
                            conflicting numbers in exec reporting. I
                            proposed and got buy-in on a single
                            definition, now used org-wide, which
                            resolved a recurring source of exec distrust
                            in our metrics."

The strong version demonstrates the staff-level pattern directly: identifying a problem that wasn't assigned, working across team boundaries without formal authority, and producing an outcome measured at the org level rather than the project level. Deliberately seeking out (or creating) opportunities that fit this pattern — rather than only executing assigned projects extremely well — is the actual lever for this transition.

A self-audit framework

import pandas as pd

self_audit = pd.DataFrame({
    "dimension": ["Technical depth", "Scope of ownership", "Cross-team influence",
                  "Mentorship/multiplier effect", "Business impact framing"],
    "current_evidence": [
        "Strong — shipped 3 production models this year",
        "Medium — owns one team's roadmap, not cross-team",
        "Weak — mostly works within own team",
        "Weak — no formal mentees, doesn't review others' work",
        "Medium — reports metrics, rarely ties to $ decisions",
    ],
    "target_for_next_level": [
        "Maintain",
        "Own a cross-team initiative",
        "Lead one initiative requiring buy-in from 2+ other teams",
        "Take on 1-2 mentees; review others' analyses regularly",
        "Reframe next quarter's work explicitly in decision/dollar terms",
    ],
})
print(self_audit.to_string(index=False))

Running this kind of audit honestly — and specifically identifying the one or two dimensions with the weakest current evidence — turns "get promoted" from a vague goal into concrete, plannable actions for the next one or two quarters, rather than something that happens passively as a side effect of doing good project work.

The generalist-vs-specialist question

Generalist path (most common at senior+):
  Broad competence across the DS stack (analysis, modeling, some
  engineering, stakeholder work). Valuable for ambiguous, cross-cutting
  problems. Ceiling: staff/principal generalist or people management.

Specialist/IC path (viable, less common):
  Deep expertise in one area (causal inference, NLP, a specific domain).
  Valuable when the org has enough volume of that specific problem type
  to justify a dedicated expert. Ceiling: distinguished/principal
  specialist, often org-wide technical authority in that narrow area.

Neither path is strictly "more senior" — the mistake is drifting into specialization by default (because it's comfortable) without checking whether the org actually has enough volume of that specific problem to sustain a specialist career track, versus deliberately choosing it because the organization's needs and your interests both point that way.

Internal vs. external growth

Signals it might be time to look externally rather than push for
internal promotion:
- The org genuinely lacks staff+ scope of problems (small company,
  thin DS function) — no amount of individual growth creates scope
  that doesn't exist
- Consistently told "not yet" without a specific, actionable gap named
- Your growth area (e.g. ML infra) isn't a direction the org is investing in

Signals internal growth is still the better bet:
- Concrete gaps were named and are addressable within 1-2 quarters
- The org is growing / has emerging scope you could plausibly claim first
- You have unusually good institutional context that resets to zero elsewhere

Cheat sheet

Level	Differentiator
Junior → Mid	Independence on well-scoped work
Mid → Senior	Owns ambiguous problems, decides what to work on
Senior → Staff	Cross-team influence without formal authority; org-level outcomes
Staff → Principal	Sets direction across the org, judged on org-wide outcomes
Any level	Multiplier effect (mentorship, reusable frameworks) compounds faster than solo output

How It Actually Works

Why staff-level work is structurally harder to attribute — the mechanism, not just the observation. A single-owner project has a one-to-one mapping between effort and outcome that's trivial to state in a promotion packet ("I built X, it did Y"). Cross-team influence work (getting three teams to converge on one metric definition) produces its outcome through other people's subsequent decisions and executions — the causal chain from "you did the convincing" to "exec reporting got more reliable" runs through several other people's independent actions, which is exactly why it reads as diffuse and is easy to under-credit even when it's the higher-leverage piece of work: leverage, by definition, means your effort is multiplied through others rather than converted 1:1 into your own output.

The multiplier effect compounds mathematically, not just rhetorically. If mentoring or a shared framework improves k other people's output by even a small factor ε each, the aggregate impact (k × ε × their individual output) scales with the number of people affected — and grows every period those people keep using what was taught or built, whereas an individual's solo output is bounded by their own single stream of hours. This is the same "why average across many uncorrelated things beats one big number" logic behind why organizations that measure staff+ impact explicitly weight reusable frameworks and mentee output alongside individually-shipped models — a multiplier applied across a growing denominator (more people, more time) mechanically outpaces a fixed numerator (one person's direct output) past some scale, even if the multiplier's individual effect looks small at any single point in time.

The self-audit table works as a gap-identification tool for the same reason a model's feature-importance or residual analysis does: it forces an explicit, itemized comparison between current state and target state across independent dimensions, rather than one global "am I doing well" impression that tends to be dominated by whichever dimension is most recently salient (usually technical depth, since it's the most legible and frequently reinforced). Breaking the assessment into separate axes is what surfaces that a dimension can be systematically neglected (cross-team influence, mentorship) precisely because it never shows up as a blocking problem in day-to-day work — only in an explicit, deliberately structured audit against the next level's actual bar.

Exercise

Fill out the self-audit table above honestly for your actual current situation (or a plausible hypothetical). Pick the single weakest dimension and write a concrete, one-quarter action plan to build evidence on it — not a skill to learn, but a specific initiative or opportunity you could pursue or create.