Google: ML Engineer & AI Engineer, India (2025)

• Recall (the share of frauds caught) usually matters more, because a missed fraud costs real money and trust.
• But precision can't collapse: false alarms block genuine customers, overload investigators and hurt the customer experience.
• Correct answer: it depends on cost.
• Estimate the cost of a false negative (the fraud amount) and of a false positive (review cost plus customer friction).
• Choose the threshold that minimises expected cost.
• Often use tiers: auto-block when the score is very high, step-up authentication (OTP) in the middle, and allow when low.
• Report PR-AUC and recall at a fixed precision (for example, recall @ 90% precision), because fraud data is highly imbalanced.

• Clarify: which buildings (malls, offices), the device (a phone, possibly with a wearable), offline needs, latency budget (under 200 ms for obstacle alerts), and privacy.
• Sensing: camera + IMU; optionally BLE beacons or Wi-Fi for positioning, or a pre-mapped building (indoor maps / visual positioning).
• On-device ML (real time):
• Obstacle and hazard detection (stairs, doors, people) with a quantized detector such as MobileNet or YOLO-nano.
• Depth estimation.
• Text and sign reading (OCR).
• Localisation and routing: visual-inertial odometry plus map matching; A* path planning on an indoor graph.
• Output: spatial audio and haptics with concise instructions ("door, 2 metres, slightly left"); a voice query for help.
• Cloud (optional): map updates and a VLM for rich scene description when the user asks.
• Evaluation: obstacle recall (safety-critical), positioning error, task completion time, and user studies with blind participants.
• Responsible AI: fail-safe behaviour on low confidence, battery use, and on-device privacy.

def coin_change(coins, amount):
    INF = amount + 1
    dp = [0] + [INF] * amount
    for a in range(1, amount + 1):
        for c in coins:
            if c <= a:
                dp[a] = min(dp[a], dp[a - c] + 1)
    return -1 if dp[amount] == INF else dp[amount]

• Bottom-up DP is O(amount × coins).
• How to talk through it: define the state (dp[a] = fewest coins for amount a), the transition, the base case and the order of filling. Start from the recursive brute force, add memoisation, then convert to a table.
• Leave 10 minutes for edge cases and complexity. This is where one candidate ran out of time.

def longest_unique(s):
    last, start, best = {}, 0, 0
    for i, ch in enumerate(s):
        if ch in last and last[ch] >= start:
            start = last[ch] + 1
        last[ch] = i
        best = max(best, i - start + 1)
    return best

• A sliding window with the last-seen index of each character: O(n) time, O(min(n, alphabet)) space.

• Metrics: PR-AUC and recall at a fixed precision; never accuracy.
• Data: stratified splits; time-based validation (fraud patterns drift).
• Techniques:
• Class weights or a focal loss.
• Undersampling the majority class.
• SMOTE, carefully and only on the training fold.
• Anomaly-detection features.
• Models: gradient boosting (XGBoost or LightGBM) with scale_pos_weight, plus graph features (shared device or IP).
• Calibration and thresholds chosen on business cost; monitoring for drift and feedback delay, because labels arrive late.

• Offline: NDCG@k, MRR and recall@k on a held-out, time-split log; check for position bias (use counterfactual or IPS estimates).
• Online:
• Shadow mode first.
• Then an A/B test with guardrail metrics (latency, complaints).
• A pre-registered primary metric (CTR or conversion) and a sample-size and power calculation.
• Watch out for:
• Novelty effects (run at least 1–2 weeks).
• Segment-level regressions.
• Feedback loops.
• Ramp up 1% → 10% → 50% → 100% with a kill switch.

• Google looks for collaboration, humility, comfort with ambiguity and user focus.
• Example: "The requirements for the internal search tool were vague. I ran five short user interviews, wrote a one-page doc with three options and their trade-offs, and built a two-day prototype for the riskiest part. The team picked option B, and adoption reached 70% of analysts in a month. I also paired with a junior colleague to hand over the indexing code."
• Show that you listen, write things down and include others.