Handle scope creep and teammate conflict

# How to Answer Effectively - Use STAR: Situation (context), Task (your goal/constraints), Actions (your decisions and why), Results (metrics, lessons). - Quantify impact: business (revenue, conversion), technical (latency, error rate), operational (incidents, on-call). - Show judgment: risk assessment, alignment, prioritization, trade-offs. --- ## Example Answer 1 — Ownership Beyond Role Situation - I was a data scientist on a recommendations team. Our online model underperformed the offline metrics. Ad-hoc checks suggested inconsistent event logging across web and mobile, but data engineering did not have bandwidth to re-instrument for ~6 weeks. Task - Ensure consistent, reliable behavioral data to unblock model iteration within the current quarter, without derailing existing roadmap commitments. Actions 1. Risk assessment - Identified risks: breaking downstream dashboards, creating duplicate events, and delaying our sprint goals. - Mitigations: backward-compatible event schema, feature flags, shadow validation, and rollback plan. 2. Stakeholder alignment - Mapped stakeholders: PM (roadmap impact), Data Eng (standards/maintenance), Mobile/Web eng (app changes), Analytics (dashboards), Marketing (attribution). - Wrote a 1-page proposal: problem, options (wait vs. patch vs. schema+pipeline), timeline, success metrics, risks, owners. 3. Manager alignment and prioritization - Proposed a time-boxed, 2-week “stabilize data plane” spike using 20% of my capacity for 6 weeks, offset by de-scoping one exploratory model experiment. Agreed on success criteria and a weekly check-in. 4. Concrete actions - Audited top 20 events; found 12% null product IDs and inconsistent timestamps (ms vs. s). - Drafted a unified event contract (naming, types, required fields, versioning) and added JSON schema validation in CI. - Partnered with mobile/web engineers to standardize client emitters; added sampling to reduce volume by 10%. - Built dbt models and an Airflow DAG to backfill 90 days, with data quality checks (null thresholds, referential integrity). - Created a comparison harness to measure offline/online feature parity and drift. - Shipped behind a flag; ran a 2-week shadow period with dashboards tracking nulls, freshness, and parity. 5. Delivery management - Maintained weekly demo to stakeholders; documented the schema and ownership handoff back to Data Eng for long-term maintenance. Results - Data quality and reliability - Reduced null product IDs from 12% to 1.1%; corrected timestamps eliminated 8% of out-of-window events. - Improved data freshness from 48 hours to 4 hours. - Model and business impact - Offline/online metric gap reduced from 6.5 pp to 1.2 pp; launched model iteration yielded +3.2 pp CTR and +0.9 pp conversion. - Estimated +$450k/quarter incremental revenue; 60% fewer analytics incident tickets. - Process impact - Established an event schema standard and validation that new features adopted. Learnings - Instrumentation and data contracts are leverage points for ML impact. - Influence without authority: circulate a concise proposal, invite feedback, and secure explicit scope/ownership. - Time-box scope and define success metrics before taking on work outside your lane to avoid becoming the permanent owner. --- ## Example Answer 2 — Conflict and Delivery Situation - On a propensity modeling project for lifecycle marketing, I advocated shipping a calibrated gradient-boosted model meeting strict latency/operability constraints. A senior ML engineer preferred a deep model with slightly higher offline AUC but higher latency and operational complexity. Tension rose as we neared a campaign deadline. Task - Resolve the disagreement quickly, select an approach that meets business and technical constraints, and deliver in time for the campaign. Root Cause - Misaligned decision criteria: one side optimized for incremental AUC; the other prioritized end-to-end delivery risk (latency, ops burden, and integration timelines). Success metrics and constraints were not explicit. Options Considered 1. Adopt the deep model; accept latency and ops complexity. 2. Ship the simpler model now; revisit the deep model later. 3. Run a time-boxed bake-off with agreed decision criteria and deploy whichever meets constraints and impact goals. De-escalation and Escalation - De-escalated by reframing around shared objectives and constraints: defined target metrics (AUC ≥ 0.72, mean inference latency ≤ 50 ms p95, cost ≤ $X/day, feature stability), and business deadlines. - Proposed a 1-week bake-off with a clear rubric; secured PM buy-in. - Escalated only to clarify resource allocation and tie-breaker expectations if the bake-off was inconclusive; documented the decision log. Keeping Delivery on Track - Parallelized work: - I productionized the GBDT with monotonic constraints and Platt scaling; set up shadow traffic in the feature store. - The engineer containerized the deep model with optimized serving using ONNX and quantization. - Built a shadow A/B harness to compare models on live traffic: tracked AUC, calibration (ECE), latency p95, and failure rate. - Prepared fallbacks: if neither met latency, default to a well-tuned logistic regression with sparse features. Outcome - Bake-off results: Deep model AUC 0.755, p95 latency 120 ms; GBDT AUC 0.743, p95 latency 38 ms, well within SLA. Calibration favored GBDT (ECE 0.03 vs. 0.07). - Shipped GBDT for the campaign; achieved +1.8 pp uplift in conversion and +$220k incremental revenue over 6 weeks. - Scheduled a post-campaign spike to optimize the deep model’s serving path. What I’d Do Differently - Establish a decision framework and constraints at project kickoff, including a tiebreaker rubric. - Use a RACI and a short PRD to align on goals, risks, and SLAs. - Pre-mortem to surface risks (latency, cost, ops) early and avoid last-minute conflict. --- ## Pitfalls to Avoid - Vague outcomes: Always include metrics and time bounds. - Ownership creep: Secure time-boxing and handoff plans when stepping outside your role. - Escalating too early: Try data-driven de-escalation first; escalate to unblock, not to win. - Ignoring constraints: Offline gains that violate latency/cost/ops are not wins. ## Quick Template You Can Reuse - Situation: One sentence on context and stakes. - Task: Your objective and constraints. - Actions: - Alignment: Stakeholders, decision doc, success metrics. - Risk & prioritization: Top risks, mitigations, time-boxing, trade-offs vs. roadmap. - Execution: 3–5 concrete steps you owned. - Results: 3–5 metrics across business, technical, and operational dimensions. - Learnings/Next time: One process improvement and one technical improvement.