Supply Chain & Logistics

A WMS uses several strategies to minimise picker travel (the largest cost in warehouse operations): 1) Wave planning — group orders going to the same carrier or zone into a 'wave' so a picker handles multiple orders in one pass. 2) Zone picking — divide warehouse into zones, each picker handles only their zone; orders assembled at pack station. 3) Batch picking — one picker collects items for N orders simultaneously using a multi-compartment cart. 4) Optimal pick path — within a zone, sort pick list by aisle-then-bin sequence (serpentine or return path depending on warehouse layout). 5) Slotting — fast-moving SKUs placed closest to pack station and at ergonomic height (golden zone). 6) Goods-to-person robotics (GreyOrange, Kiva) — robots bring shelves to stationary pickers, eliminating travel entirely. Modern WMS also uses ML to predict which orders will arrive together and pre-positions pickers.

VRP asks: given N delivery locations and K vehicles at a depot, find the optimal set of routes minimising total distance/time. It is NP-hard (no polynomial-time exact solution). Practical approaches: 1) Clarke-Wright Savings algorithm — merge routes that share a common savings value. 2) Genetic algorithms / simulated annealing — metaheuristic search. 3) Google OR-Tools — open-source combinatorial optimisation library widely used in industry. Real-world constraints add complexity: time windows (deliver between 10am–12pm), vehicle capacity (weight/volume), driver hours (labour law), traffic conditions (Google Maps API live data), priority orders (same-day vs next-day). Companies like Locus, FarEye, and Delhivery's internal TMS solve variants of VRP in near-real-time (seconds) for thousands of shipments using a combination of heuristics and ML-based warm starting.

Design components: 1) Event ingestion — scanning events pushed from barcode scanners at hubs, GPS pings from vehicles, DE app location updates. Use Kafka for high-throughput event streaming (millions of events/hour). 2) Event processing — stream processor (Kafka Streams / Flink) normalises events, resolves to shipment ID, applies milestone state machine (Picked Up → In Transit → At Hub → Out for Delivery → Delivered). 3) Storage — current status in Redis (O(1) lookup by AWB), full event history in Cassandra (append-only time-series). 4) Customer-facing tracking — React page polling /api/track/{awb} or WebSocket push. 5) ETA prediction — ML model taking current location, historical delivery time for that route, time-of-day, traffic — served via a low-latency feature store. 6) Exception detection — stream analytics job detects SLA breaches (no scan event in X hours), triggers alert via webhook to merchant. Scale challenge: Delhivery handles 1M+ active shipments daily — partition Kafka by AWB hash, shard Redis by AWB range.

OTIF — On-Time In-Full — measures the percentage of orders delivered on time AND with the correct items and quantity. It is the single most critical logistics KPI because: 1) 'On Time' — delivery within promised SLA window; late delivery = poor customer experience, FMCG retailers penalise suppliers (Walmart charges a fine for OTIF breach). 2) 'In Full' — complete order without short-shipments or substitutions; partial delivery = incorrect inventory at customer, order re-processing cost. Calculation: OTIF % = (Orders delivered on time AND in full) / Total orders × 100. Industry benchmarks: B2C e-commerce: 95%+ OTIF expected; FMCG modern trade: 97%+; quick commerce: 99%+. Root causes of OTIF failure: inventory stockouts, warehouse processing delays, carrier failures, address quality issues, route planning errors. WMS, TMS, and demand planning systems all directly contribute to improving OTIF.

Demand forecasting predicts future sales to drive inventory and replenishment decisions. Approaches by complexity: 1) Statistical models (baseline): Moving Average — average of last N weeks; Exponential Smoothing (Holt-Winters) — weighted average giving more weight to recent data, handles seasonality and trend. 2) Causal models: Add external factors — promotions, price changes, weather, events (e.g., Diwali spike). Regression-based. 3) ML models (modern): XGBoost / LightGBM — tabular features: historical sales, price, promotions, day-of-week, holidays, competitor events. 4) Deep learning: LSTM / Temporal Fusion Transformer — for complex multi-variate time series. Challenges: new product introduction (no history), long-tail SKUs (sparse data), cannibalization (new product eats existing SKU sales), promotional uplift quantification. In practice, most large retailers use a hierarchy: ML model for top SKUs, statistical for mid-tail, min-max reorder for long-tail.