Intuit (Mountain View) — AI Engineer Fintech Case Interview

This case mirrors Intuit’s customer-obsessed, data-driven, and compliance-first interview style. You will design an end‑to‑end AI capability for one of Intuit’s flagship products under real-world constraints. Typical prompts: (a) QuickBooks: auto‑categorize bank transactions and generate anomaly/spend alerts; (b) TurboTax: a conversational filing assistant that ingests W‑2/1099s and explains deductions; (c) Credit Karma: a ranking system for credit-card recommendations with reason codes. What the interviewer evaluates: 1) Problem framing with deep customer empathy—clarify target users (e.g., small-business owners, filers, tax pros), success criteria, and edge cases; articulate jobs-to-be-done and how the solution reduces time-to-value. 2) Data and privacy: identify sources (bank feeds, receipts, tax forms, clickstream), schema, data quality, labeling strategy, and governance; handle PII/PCI and tax data with least-privilege access, tokenization, and purpose limitation; discuss retention, consent, and auditability. 3) Modeling strategy: choose between classical ML (e.g., gradient boosting for categorization/fraud) and LLM/RAG for understanding documents and generating explanations; cover feature engineering, embeddings, prompt design, retrieval, safety filters, and human-in-the-loop review for uncertain predictions. 4) System design for peak scale and low latency: propose offline training pipeline, online feature store, model registry, canary/shadow deployments, and autoscaling to handle tax‑season spikes; define SLOs (e.g., P95 <150 ms for ranking; <500 ms for LLM responses with streaming), fallback paths, and cost controls. 5) Measurement and experimentation: define offline metrics (precision/recall, AUC, calibration, abstention rate), online KPIs (task completion time, auto‑categorization coverage, acceptance rate, lift in TPV or conversion), guardrail metrics (complaint rate, safety flags, latency, fraud loss), and an A/B plan with ramp criteria and segment analysis. 6) Responsible AI in a regulated domain: fairness checks (ECOA/FCRA considerations), explainability (reason codes, model cards), red-teaming for prompt injection/data leakage, and incident response. 7) Communication and collaboration: crisp tradeoffs, executive summary, and how you would partner with design, product, compliance, and tax experts. Format and flow (typical): 5 min clarifying questions and success metrics; 20 min data and modeling approach; 20 min system/serving design; 10 min metrics/experiments; 10 min risks, safety, and rollout. Expect to sketch APIs, data contracts, and an evaluation plan, and to defend tradeoffs under time and compliance constraints.