Tesla Data Analyst Case Interview: Fleet Telemetry Reliability and Yield Optimization

Purpose: A hands-on, numbers-first case reflecting Tesla’s bias for rapid, first-principles problem solving and real-world impact. You will analyze messy, high-volume data to diagnose an issue, propose an experiment, and quantify business and sustainability outcomes. Context (aligned to Tesla’s mission): A recent OTA firmware change aimed at improving thermal efficiency in hot climates coincides with an uptick in drivetrain power-derate events on a subset of vehicles. Separately, Gigafactory reports a small dip in cell line yield on packs using the new firmware. Your task is to determine what’s happening, whether it’s causal, and what to do next—quickly and pragmatically. Format & timing (~70 minutes): - 5 min: Problem brief and clarifying questions (expect direct, concise back-and-forth; interviewers may challenge assumptions). - 35 min: Hands-on analysis in a shared SQL/Python environment (Pandas/NumPy allowed). Speed over polish; focus on correctness, reproducibility, and crisp reasoning. - 10 min: Experiment design (A/B or phased rollout) with safety/quality guardrails. - 15 min: Executive readout: a whiteboard-style, numbers-first summary (no slideware required) with a clear recommendation and next steps. - 5 min: Culture fit probe: ownership mindset, ability to dive deep while staying scrappy, and willingness to challenge constraints. Data provided (representative, intentionally imperfect): - fleet_telemetry (partitioned by date): vehicle_id, ts, region, ambient_temp_c, inverter_temp_c, coolant_temp_c, power_kw, speed_kph, soc_pct, pack_revision, firmware_version, derate_flag, odometer_km. - firmware_rollout_log: vehicle_id, firmware_version, rollout_ts, cohort (pilot/control), region, vehicle_config. - mfg_cell_yield_daily: date, line_id, pack_revision, yield_pct, scrap_rate, top_scrap_reason. What you’ll do: 1) SQL diagnostic: Compute derate_rate per 10k miles by firmware_version, ambient temperature bin, and pack_revision. Join rollout logs; identify segments with statistically significant increases vs. pre-rollout baselines. Show exact queries and intermediate sanity checks (row counts, nulls, unit consistency). 2) Root cause framing: Apply first-principles reasoning to hypothesize mechanisms (e.g., tighter thermal limits causing protective derates under high ambient + certain pack revisions). Use simple plots or aggregated tables; prioritize explainability. 3) Quick model/rule: In Python, prototype a lightweight threshold rule or logistic regression to predict derate_flag using a few interpretable features. Report precision/recall and a confusion matrix on a holdout split. Emphasize false-negative risk (safety) and operational interpretability. 4) Experiment design: Propose a phased OTA mitigation (e.g., reverting a parameter for hot-climate cohorts). Define primary metric (derate_rate), guardrails (thermal headroom, warranty incidents, energy consumption), sample size/time-to-detect, and a clear rollback plan. 5) Impact sizing: Back-of-the-envelope ROI. Translate reduction in derates into avoided service visits/warranty cost, kWh saved, and estimated CO2e avoided for grid-charged energy. Call out assumptions and sensitivity. 6) Recommendation & next steps: A single-page (or whiteboard) summary with: decision, rationale, risks, instrumentation/telemetry gaps to close, and owners/timelines. Evaluation rubric (Tesla-specific): - Data rigor (35%): Trust-but-verify of joins, handling of missing/erroneous telemetry, unit checks, and baseline comparisons. - Practical SQL/Python (30%): Clear, performant queries on large partitions; readable, reproducible notebooks; preference for simple, robust methods over complex models. - Decision quality (20%): Sound experimental design, explicit trade-offs (safety, cost, sustainability), and an action-oriented recommendation. - Communication (10%): Concise, numbers-first storytelling; anticipates pushback; minimal fluff. - Culture alignment (5%): Ownership, bias for action, willingness to challenge constraints, and mission focus on accelerating sustainable energy. Constraints & expectations: - Assume data is imperfect; document the minimum viable checks you’d ship today and what you’d harden next week. - No heavy slideware; whiteboard or a brief written summary preferred. - Be explicit about safety and quality guardrails before proposing changes. - Show work: paste core SQL, describe feature choices, and state all assumptions with impacts. Common follow-ups from interviewers: - How would you monitor this post-rollout in near-real-time (dashboards, alert thresholds, on-call signals)? - What additional logs/telemetry would you add to isolate causality? - How does your recommendation change for cold climates, different pack revisions, or Supercharger-heavy duty cycles? - If manufacturing yield is the true bottleneck, how would you adapt your analysis to isolate process steps driving scrap?