Assess Amazon Leadership Principles in Behavioral Interviews

## How to approach the LP round - Use **STAR + L**: Situation (context, scope, stakes), Task (your goal and constraints), Action (3–5 specific decisions, alternatives considered, trade-offs, mechanisms), Result (quantified customer/business impact), and Learnings (what you institutionalized — dashboards, SOPs, guardrails, LPs demonstrated). - **Quantify** with concrete metrics (e.g., +5.2% conversion, −35% latency, $4.2M ARR, p95 84ms, 6 hours/week saved). If numbers are sensitive, use relative terms ("~20% lift"). - **Name the Leadership Principles** you exemplified. At **L5** scope, show cross-functional influence, end-to-end accountability, comfort with ambiguity, measurable business outcomes, and mechanisms that scale. - **Prep:** draft 6–8 stories ahead of time, map each to 2–3 LPs, and be ready to **Dive Deep** technically (data, model choices, trade-offs). ### LP mapping cheatsheet (use selectively) - Conflict resolution → Earn Trust, Dive Deep, Customer Obsession, Ownership - Tight timeline → Bias for Action, Deliver Results, Invent and Simplify - Improve processes → Insist on the Highest Standards, Dive Deep, Ownership - Disagree and Commit → Have Backbone; Disagree and Commit, Are Right, A Lot - High standards → Insist on the Highest Standards, Dive Deep - Invent and Simplify → Invent and Simplify, Think Big - Think Big → Think Big, Customer Obsession, Ownership - Exceed expectations → Deliver Results, Ownership, Bias for Action - Missed customer expectations → Customer Obsession, Dive Deep, Bias for Action - Act fast / limited data → Bias for Action, Are Right, A Lot, Dive Deep - End-to-end ownership → Ownership, Insist on the Highest Standards, Deliver Results --- ## Data-science-tailored STAR examples (adapt to your own stories) ### 1) Conflict with a colleague (metric disagreement) - **S:** A PM wanted to judge a recommender by CTR; I believed revenue per session was the right North Star for our retail surface. - **T:** Align on a success metric before launch. - **A:** Pulled 6 months of data showing high-CTR items cannibalized higher-margin items; ran an offline replay and a 1-week A/A; proposed a weighted composite (revenue + attach rate, bounded by dwell time); facilitated a 30-minute decision review. - **R:** Adopted the composite metric; revenue per session +8.4%, CTR +1.1% (vs. the alternative's +4% CTR but −1.9% revenue), returns −6%. - **LPs:** Dive Deep, Customer Obsession, Earn Trust, Are Right, A Lot. ### 2) Very tight timeline (10-day MVP) - **S:** VP demo in 2 weeks for a personalized deals page. - **T:** Ship an MVP safely and on time. - **A:** MoSCoW-scoped features; reused the existing feature store; trained a baseline gradient-boosting model; set guardrails (p50 latency <80ms, null-safe fallbacks); parallelized with clear DRIs. - **R:** Shipped in 10 days; A/B showed +3.2% CTR, +1.4% revenue per visit; follow-on hardening cut latency 35% in 3 weeks. - **LPs:** Bias for Action, Deliver Results, Invent and Simplify, Ownership. ### 3) Improved an existing process (reporting automation) - **S:** A weekly KPI deck took 6 analyst-hours and had frequent inconsistencies. - **T:** Improve accuracy and cut cycle time. - **A:** Centralized definitions in dbt; added data tests (freshness, uniqueness); automated the pipeline in Airflow; published a Looker dashboard; wrote a data dictionary. - **R:** Manual time −90% (6h → <30m), data issues −85%, on-time delivery 100% (~300 analyst-hours/year saved). - **LPs:** Insist on the Highest Standards, Dive Deep, Ownership. ### 4) Have Backbone; Disagree and Commit (then execute with excellence) - **S:** Leadership chose a heuristic rule-based pricing update over my proposed demand-elasticity model ahead of a seasonal spike. - **T:** Voice the risk (margin dilution, customer fairness), align on evaluation, and — if the decision stands — execute flawlessly. - **A:** Presented a pre-mortem modeling −1% to −3% margin risk under inventory constraints; pre-aligned success metrics and guardrails (no price change >8% without approval). The decision stood, so I committed: productionized the heuristic with comprehensive logging and a 20% holdout for honest evaluation. - **R:** After 2 weeks: +0.7% revenue but flat margin and acceptance −1.2pp. With the trust built and the holdout data, I got the green light to A/B the elasticity model on underperforming segments → +2.9% revenue, +1.1pp margin, then a broad rollout. - **LPs:** Have Backbone; Disagree and Commit, Are Right, A Lot, Ownership. (Separating advocacy from execution sustains both velocity and trust.) ### 5) Insist on the highest standards (caught data leakage) - **S:** A churn model's offline AUC was 0.84 but online lift was negligible. - **T:** Ensure genuine model quality before re-launch. - **A:** Added feature-recency unit tests, leakage checks, and schema versioning; rewrote cross-validation to be time-based; blocked launch until resolved. - **R:** Found leakage in a late-arriving refund feature; retrained to an honest AUC 0.78; post-fix retention lift +3.9% vs. +0.4% prior. - **LPs:** Insist on the Highest Standards, Dive Deep, Customer Obsession. ### 6) Invent and Simplify (shared feature platform) - **S:** Multiple teams re-implemented the same features, causing drift and duplicated cost. - **T:** Create a simple, shared path to production features. - **A:** Built a lightweight feature registry with checks, lineage, and backfills; authored templates and a contribution guide. - **R:** Duplicate features −60%; model onboarding 4 weeks → 1.5; infra cost −20% via reuse. - **LPs:** Invent and Simplify, Ownership, Frugality. ### 7) Think Big (causal, long-term measurement) - **S:** Teams optimized local CTR; the org lacked causal, long-term impact measurement. - **T:** Elevate measurement to long-term customer value. - **A:** Piloted an uplift-modeling + sequential-testing framework with long-term holdouts; created a central "customer value" metric. - **R:** Within two quarters, 4 launches shifted from CTR-first to value-first → +2.1% 90-day revenue with neutral engagement; the platform was adopted by 3 orgs. - **LPs:** Think Big, Customer Obsession, Are Right, A Lot. ### 8) Exceeded expectations (A → A, B, C) - **S:** Asked to deliver a propensity model to prioritize sales leads. - **T:** Deliver the model (A). - **A:** Shipped the model plus (B) a self-serve dashboard with cohort drill-downs and (C) an SDR outreach-cadence playbook; added SHAP-based explainability. - **R:** Conversion +14% in pilot; SDR ramp time −25%; leadership rolled it out org-wide. - **LPs:** Deliver Results, Ownership, Bias for Action. ### 9) Customer Obsession — missed expectations, recovery, and do-differently - **S:** A personalized-recommendations launch on a high-traffic page drew complaints within 48 hours; CTR dropped 5.6% and contact rate rose 18%. - **T:** Stabilize the customer experience within 72 hours without abandoning the personalization roadmap. - **A:** Rolled 20% of traffic back to a popular-items fallback; ran a rapid RCA (cohort CTR, error logs, feature-completeness) and found 14% of items missing key attributes; hotfixed feature imputation; added data-quality monitors (missingness alert >2%) and a canary rollout (5% → 25% → 50% → 100%); tagged 200 complaint tickets to classify failure modes. - **R:** CTR recovered to +2.3% above baseline within a week; contact rate −22% below baseline; data missingness sustained <0.5%; ~$1.1M incremental quarterly revenue. - **Do differently:** phased rollout by segment with explicit guardrails (SRM checks, p95 latency <150ms); pre-launch dogfood and synthetic cold-start tests; define customer-harm leading indicators as automatic kill-switches. - **LPs:** Customer Obsession, Dive Deep, Bias for Action, Insist on the Highest Standards. ### 10) Bias for Action + Are Right, A Lot — acted fast with limited data - **S:** A surge of fraudulent sign-ups began abusing promo credits; labels were sparse and finance projected $250k weekly exposure. - **T:** Cut losses within 48 hours with minimal customer friction and low false positives. - **A:** Triangulated signals (device-fingerprint entropy, signup velocity per IP/BIN, unsupervised anomaly scores); pulled a stratified sample of 100 accounts for manual review to estimate a ~28% baseline fraud rate (±8–10% given the sample size); deployed a lightweight rules-plus-score threshold with a human-review queue for ambiguous cases, plus a rollback switch and daily calibration; monitored chargebacks, appeal rate, conversion, and cohort LTV. - **R:** Fraud loss −76% in 72 hours; false-positive rate held <2.5% (target <3%); legitimate conversion dipped 0.6pp for 3 days then normalized after threshold tuning. - **How I stayed "right, a lot" under uncertainty:** chose conservative thresholds via confidence bounds and sensitivity checks; used a human-in-the-loop queue to cap harm while learning; logged features/decisions and backtested weekly as labels accrued. - **LPs:** Bias for Action, Are Right, A Lot, Dive Deep. ### 11) Ownership + Insist on the Highest Standards — end-to-end delivery under pressure - **S:** Churn rose in a B2C subscription product; leadership wanted a retention uplift within a quarter. - **T:** Own an end-to-end churn-prediction and intervention system (data pipeline → model → orchestration → measurement) on a 10-week deadline. - **A:** Built a feature pipeline (events, support tickets, payment signals) with SLAs, unit tests, and p95 freshness monitoring; trained calibrated gradient boosting with monotonic constraints and cost-sensitive thresholds per segment; pre-registered metrics and ran a powered 50/50 RCT across 1.2M users with SRM and CUPED variance reduction; triggered tiered offers behind a p95 inference latency <100ms (batch + online cache), shadow-tested before full enablement; added integration tests, drift alerts, and a fairness/compliance red-team review. - **R:** 60-day churn −3.8pp (22.1% → 18.3%), +$4.2M ARR; p95 latency 84ms; alert-driven ops cut incident MTTR by 60%; shipped on time with mechanisms still in place. - **LPs:** Ownership, Insist on the Highest Standards, Deliver Results, Dive Deep. --- ## Preparation checklist - Draft 6–8 STAR stories; map each to 2–3 LPs; cover your role, scale, metrics, and customer impact. - Bring numbers: n users, $ impact, % change, latency, error rate, precision/recall, p90/p99, cost-to-serve. - Make trade-offs explicit (speed vs. quality, precision vs. recall, latency vs. cost). - Name mechanisms: canary, SRM checks, A/A tests, power analysis, guardrails, alerts, kill-switches, rollbacks. - Practice 2-minute summaries with optional 1–2 minute technical deep-dives. ## Common pitfalls to avoid - Vague or unverified results — always quantify against a baseline; use proxy metrics if needed. - Team-only credit — articulate your unique decisions and actions. - No customer link — tie every result to customer/business outcomes. - Over-indexing on models vs. customer impact and durable mechanisms. - Blaming without owning, or failing to reflect on what you'd change next time. ## Quick guardrails for experimentation stories - **Before:** define success metrics and guardrails; run a power analysis; plan an A/A test. - **During:** monitor SRM, leading indicators, and error budgets; keep kill-switches enabled. - **After:** validate uplift with confidence intervals; segment for heterogeneity; run a post-mortem and turn learnings into mechanisms. ## Mini STAR note template (copy/paste) - **Title:** [e.g., "Automated KPI pipeline"] - **Situation / Task / Action / Result / Learnings** - **LPs:** [2–3 relevant principles]

Rubric: this is the Amazon Data Scientist (L5) Leadership Principles behavioral round, not a single question — the candidate must field a battery of 'Tell me about a time…' prompts spanning conflict, tight timelines, process improvement, Disagree and Commit, high standards, Invent and Simplify, Think Big, exceeding expectations, missed customer expectations, acting on limited data, and end-to-end ownership under pressure. Strong answers use STAR+Learnings, quantify against a baseline, make trade-offs and mechanisms explicit, and explicitly map to 2–3 Leadership Principles each. At L5 the bar is cross-functional influence, end-to-end accountability, comfort with ambiguity, and measurable business outcomes. Bonus signal: data-science-specific rigor (A/A and A/B tests, power analysis, SRM, CUPED, calibration, leakage checks, drift monitoring, kill-switches/rollbacks) and a genuine 'do differently' reflection.