Uber: Machine Learning Engineer

• Label: accept/decline (or timeout) per request. Mind the selection bias: only dispatched offers are logged.
• Features:
• Pickup distance and ETA; trip length and direction (towards home?).
• Surge level; time and day.
• The driver's acceptance history, hours online, earnings so far.
• Rider rating; pickup area type (airport).
• Model: gradient boosting (or a two-tower model for driver–request matching), calibrated probabilities.
• Use: the dispatch optimiser maximises expected matches = Σ P(accept) × value; avoid over-offering to the same drivers.
• Evaluation: log-loss and AUC offline; online A/B on match rate, pickup ETA, cancellation, driver earnings fairness.

• Accuracy: MAE/MAPE against actual trip times (delayed labels arrive after the trip) by city, hour, distance band; the share of trips more than X minutes late.
• Input drift: distributions of traffic, GPS quality, new road closures; prediction drift against the previous model.
• System: p99 latency, error rate, fallback rate (to a heuristic), feature staleness.
• Business: cancellations, rider complaints, driver idle time.
• Alerts and rollback: shadow first, then canary; automatic rollback if MAE regresses by more than a threshold.

• Randomness: weight initialisation, data shuffling, dropout, bagging/feature subsampling, train/test split seeds.
• Non-determinism: GPU kernels (atomic adds), multithreading, floating-point order of operations.
• Different hyperparameters or library versions; data-order effects with early stopping.
• Fixes: set seeds, use deterministic flags, pin versions; report mean ± std over several seeds.

import numpy as np
def fit_line(x, y, lr=0.01, epochs=2000):
    w = b = 0.0; n = len(x)
    for _ in range(epochs):
        y_hat = w * x + b
        dw = (2 / n) * np.dot(y_hat - y, x); db = (2 / n) * np.sum(y_hat - y)
        w -= lr * dw; b -= lr * db
    return w, b
x = np.array([1, 2, 3, 4.]); y = np.array([3, 5, 7, 9.])
print(fit_line(x, y))   # ≈ (2, 1)

• Discuss the learning rate choice, feature scaling and convergence checks (when the loss stops improving).

• The problem: interference. Treating some riders changes the drivers available to control riders, which biases user-level A/B tests.
• Designs:
• Switchback tests: randomise by city-region × time window.
• Geo/cluster randomisation.
• Driver-side vs rider-side randomisation depending on the feature.
• Metrics: marketplace health (match rate, wait time, surge frequency, earnings per hour, cancellations) plus guardrails.
• Analysis: account for carry-over between windows, use cluster-robust standard errors, run long enough to cover weekly cycles.

Answer (STAR):

• Situation: "A new pricing model inflated fares in one city after a holiday-calendar bug."
• Action: "I detected it from an alert on fare distribution, rolled back within 20 minutes, quantified the affected trips, worked with ops on refunds, and added a calendar unit test plus a canary gate."
• Result: "There were no repeat incidents, and the postmortem became the team's release checklist."
• Show speed, ownership, transparency and prevention.