Build an ETA Prediction Model for Zomato

Problem Build a model that predicts ETA (estimated time of arrival) for a food-delivery order — from order placement through to hand-off at the customer's door.

Functional requirements

• Return an ETA at checkout, before the order is placed.
• Update the ETA as the order progresses through prep, pickup, and transit.
• Decompose the estimate into prep time, rider assignment/wait, and travel time.
• Handle cold-start: new restaurants, new delivery areas, and riders with little history.

Non-functional requirements

• ~5-10k prediction requests/sec at peak meal hours; tens of millions of orders/day.
• p99 inference latency under ~50ms — the ETA blocks the checkout screen.
• Retraining daily to weekly; traffic and demand patterns shift seasonally and around events.
• Every feature must be available in real time at request time, with a defined fallback when one is stale or missing.

Key components

• Data collection: historical order timelines with per-stage timestamps (placed, accepted, prepped, picked up, delivered) as ground-truth labels.
• Feature engineering: haversine and road-network distance, historical durations on the same route, time of day, day of week, weather, live traffic/congestion, restaurant prep-time history and current kitchen backlog, rider availability and load in the area.
• Model: gradient-boosted trees as a strong baseline, or a hybrid that starts from a physical distance/speed prior and learns an ML correction on top; per-stage models summed vs. one end-to-end regressor.
• Feature store serving batch aggregates and real-time signals consistently between training and inference.
• Monitoring: predicted vs. actual drift, per-city and per-restaurant error breakdown, alerting on systematic bias.

Deep dives / trade-offs

• Loss and calibration: MAE/RMSE hide the asymmetry — being 10 minutes late costs far more than being 10 minutes early. Quantile regression to deliberately over-predict, and choosing the quantile as a product decision.
• Feedback loop: the ETA shown to the customer changes their behaviour and the rider's, so logged actuals are not a clean sample of the counterfactual.
• Cold-start hierarchy: backing off from restaurant-level to cuisine/area-level to city-level priors when history is thin.
• Train/serve skew: batch features computed over full history vs. what is genuinely knowable at request time, and leakage from post-hoc fields.