# 1) Why streaming now, and why this team? (Motivation + Fit) - Streaming is at an inflection: ad‑tiers, bundling, sports/live, and password‑sharing changes are reshaping growth and retention. That creates high‑leverage DS work in discovery/personalization, pricing, and churn prevention. - User problems I want to work on: - Discovery friction: too much choice; cold‑start for new titles and new users; multi‑profile households. - Retention: gaps between tentpoles; completion drop‑offs; re‑engagement. - QoE → satisfaction: buffering, startup latency, device fragmentation impacting play starts & completion. - How my background maps: - Personalization/Ranking: Built a home‑row ranker using embeddings + calibrated CTR models; improved session watch starts +6% with diversity constraints to avoid over‑concentration on blockbusters. - Causal inference & experimentation: Designed guardrails (play failure rate, buffering ratio, cancel intent clicks) and CUPED to shrink variance; reduced experiment runtime 25%. - Retention & LTV: Deployed churn propensity + save‑offer policies; reduced 90‑day churn 1.5pp on an ad‑supported product; LTV uplift modeling for promo targeting. - Data productization: Partnered with content & growth to translate models into weekly greenlight and promo decisions with decision thresholds and error budgets. - Why this team: The mandate spans recommendations + retention, where DS can measurably move hours‑watched, D30 retention, and ARPU. The catalog mixes prestige series, tentpoles, and unscripted—ideal for context‑aware recsys (franchise graphs, episode sequencing, and binge vs. appointment viewing). # 2) Disney+ vs. Netflix; Risks/Opportunities for HBO Max - Content strategy - Netflix: Broad four‑quadrant slate; heavy local originals (Korea, Japan, EMEA), volume to fill daily demand; strong kids + reality + crime; binge and weekly hybrids. - Disney+: Franchise/tentpole heavy (Marvel, Star Wars, Pixar), family‑safe; lower volume, high IP leverage; integrating Hulu expands into adult genres. - Pricing - Netflix: Multi‑tier incl. ads; frequent price optimization; paid‑sharing program increased ARPU; limited long‑term discounts. - Disney+: Aggressive ad‑tier pricing; frequent increases; cross‑subsidized via bundles. - Bundling - Netflix: Primarily telco/device bundles; less cross‑brand bundling. - Disney+: Strong native bundle (Disney+ / Hulu / ESPN+), and app integrations; cross‑promo leverage is high. - International growth - Netflix: Early mover; robust payments, telco partnerships, and local‑language content; deep catalogs per region. - Disney+: Rapid early expansion; volatility where sports rights (e.g., cricket) shift; improving local originals. - Data/ML maturity - Netflix: Industry benchmark—mature personalization, experimentation at scale, QoE optimization, near‑real‑time decisioning, strong semantic layers. - Disney+: Advancing via Hulu heritage and ad tech; still consolidating across brands and platforms. - HBO Max risks - Brand stretching (prestige HBO to broader Max) risks alienating core users and confusing positioning. - Content removals and price increases can spike churn; tentpole cadence gaps create re‑sub behavior. - Integration complexity (post‑merger tooling, taxonomy, metrics) can slow decisioning; debt pressure may constrain content bets. - HBO Max opportunities - Combine high‑affinity HBO series with Discovery’s broad, evergreen unscripted for stable hours‑watched and retention. - Ads tier growth: premium environment; uplift ARPU via relevant, low‑frequency ads with strong QoE controls. - Sports/live (where applicable) for appointment viewing; cross‑promo from live to on‑demand. - Personalization that understands franchises and session intent (short‑form reality vs. long‑form drama), plus lifecycle‑aware retention offers. # 3) Merger impacts on the data org + 90‑day harmonization plan A. Likely second‑order impacts - Metrics and definitions - Conflicting definitions of MAU/WAU/DAU; what counts as a "view" (e.g., 2s vs. 30s threshold); completion, sessionization, household vs. profile unit of analysis; time zones and backfill alignment. - Experimentation - Two randomization/bucketing systems; identity conflicts (account vs. device), elevated SRM, cross‑test interference, heterogeneous guardrails, and different ethics/privacy rules. - Taxonomy and metadata - Title/episode IDs, franchise hierarchies, genre tags, maturity ratings, language/territory mappings, device taxonomy; inconsistent content graphs hinder recsys and analytics. - Privacy/compliance - Divergent consent flows, retention periods, DSR processes, GDPR/CCPA interpretations, cross‑border data transfer rules. - Infra and cost - Duplicate pipelines/warehouses/BI tools; inconsistent semantic layers; monitoring gaps increase incidents and MTTR. - Org/process - Overlapping analytics teams; unclear ownership; fragmented roadmaps; slowed time‑to‑insight. B. 90‑day plan to harmonize metrics and experimentation - Guiding principles - Minimize decision risk; ship a compatibility layer first; prove parity with AA tests; document everything; privacy by design. Phase 0–15 days: Discover, freeze risk, align - Appoint executive sponsors and a Metrics & Experimentation Council (data science, analytics engineering, product, legal/privacy). - Change freeze on new global experiments; allow low‑risk, within‑platform tests only. - Inventory - Metrics: exact formulas, windows, thresholds, units (household/profile), time zones. - Experiment stack: bucketing, exposure logging, CUPED usage, sequential testing, guardrails. - Taxonomies: ID graphs (user, device, title), genre, ratings. - Privacy: consent schemas, DSR SLAs, data retention. - Define north‑star and guardrails candidates: D30 retention, hours‑watched/user, play start rate, time‑to‑first‑frame (TTFF), buffering ratio, crash rate, cancel‑intent clicks, CS contacts/1k users. Phase 16–30 days: Draft standards and a compatibility layer - Publish canonical metric specs (plain language + SQL/pseudo): - Example: A "view" = playtime ≥ 120 seconds on a unique profile within a 24‑hour window, excluding autoplay previews; parameterize for region exceptions. - Sessionization: 30‑minute inactivity timeout; tie session to profile + device. - Identity and bucketing policy: - Primary unit: profile_id; fallback: account_id; device_id only for signed‑out surfaces. - Single global hash for experiment assignment; namespace to avoid collisions. - Experiment guardrails standard: - Required: SRM check, power analysis before launch, sequential monitoring with alpha‑spending, CUPED or covariate adjustment, pre‑registered hypotheses, stop/extend rules. - Global do‑not‑exceed thresholds: e.g., play failure +0.3pp, buffering ratio +5%, cancel‑intent +0.2pp. - Build a thin semantic layer/metrics registry (YAML + validation) to map old → canonical fields. Phase 31–60 days: Pilot and harden - Backfill 6–12 months of canonical metrics; reconcile deltas; publish a parity report. - Run AA tests in top markets to validate bucketing and SRM; target SRM false positives <1% per week. - Release a sample size calculator (binary and continuous outcomes): - n per arm ≈ 2 * (Z_{1-α/2} + Z_{1-β})^2 * σ^2 / δ^2 (continuous) - For proportions: n per arm ≈ 2 * (Z_{1-α/2} + Z_{1-β})^2 * p(1−p) / δ^2 - Implement automated experiment QA: exposure logging, novelty effects checks, metric drift detection. - Taxonomy mapping tables: title_id and franchise graph reconciliation; device and geo normalization. - Privacy: unify consent flags in the identity graph; document data‑minimization and retention. Phase 61–90 days: Migrate, deprecate, govern - Migrate top 10 exec dashboards to canonical metrics; add confidence bands and annotations for definition changes. - Deprecate duplicate metrics with sunset dates; 301‑style redirects in BI. - Roll out a self‑serve experiment UI with guardrails enforced; require pre‑launch checklist. - Establish ongoing governance: monthly Metrics Council; change‑request process; documentation in a central handbook; incident post‑mortems. # 4) Cross‑office delivery with communication barriers (STAR) - Situation: Led a cross‑region launch of a recommendations model across U.S., Poland, and India teams; heavy accents and different communication norms created misunderstandings on requirements and experiment setup. - Task: Ship the model and an A/B test without QoE regressions; align DS, BE, and PM. - Actions: - Pre‑reads and visuals: Sent 1‑page briefs and sequence diagrams 24 hours before calls; used screenshots and example payloads. - Shared glossary: Defined terms like "view," "play start," "session"; posted in Confluence. - Confirmation loops: Ended meetings with a written summary of decisions, owners, and dates; asked each lead to restate their piece. - Psychological safety: Opened with a norms slide (no‑interruptions, ask clarifying questions); used anonymous Q&A in Slido; rotated facilitators. - Async first: Recorded calls; used Slack threads with action labels [DECISION], [BLOCKER]; scheduled quiet hours overlap. - Pairing: Paired DS with local engineers for code reviews; used checklists for experiment launch. - Results: Launched on time; +4.8% increase in first‑session watch starts; no guardrail breaches; cross‑team survey showed higher clarity and inclusion. # 5) Handling an "ASAP" dashboard request (clarify, define, and de‑scope with trust) - Clarify the decision and timeframe - Prompt: "What decision will this dashboard inform? By when? What action will you take if metric X moves up/down?" - Example: "We must decide in 2 weeks whether to expand the ad‑tier rollout to 3 new markets." - Define success metrics and thresholds - Primary: Ad‑tier signup share, ARPU, play start rate, TTFF, churn. - Leading indicators: Ad error rate, buffering ratio, completion rate. - Thresholds: e.g., churn must not increase >0.3pp; TTFF ≤ 2.5s p95. - Users and cadence - Who will read it? Exec vs. PM vs. Ops; frequency (daily vs. weekly) → dictates level of detail. - Scope a thin‑slice v0, then v1/v2 - v0 (48 hours): 3–5 tiles with primary KPIs and guardrails, segmented by market/device; data quality checks. - v1 (1–2 weeks): Cohorts, trends, drill‑downs, annotations. - v2: Alerts, experimentation overlays, LTV views. - Push back with options, not "no" - "We can deliver v0 by Friday covering the decision; deeper cuts next week. If we add 10 more slices now, reliability will drop; which two are must‑have?" - Guardrails and validation - Include data freshness, anomaly flags, and metric definitions inline; run parallel checks vs. existing reports for a week. - Close the loop - After delivery, instrument dashboard usage; if engagement is low, schedule a 15‑minute retraining or retire. Notes and pitfalls - Avoid vanity metrics (raw signups) without denominators (rate per eligible users). - Align metric definitions with the canonical registry; annotate any deviations. - Prefer comparatives and targets (Vs. last week; target bands) over absolute numbers only.