How Amazon, Coupang, and Walmart Use AI — And What It Means for Your Interview

Most interview prep focuses on what questions companies ask. This piece goes deeper: the AI systems Amazon, Coupang, and Walmart actually run, why they built them the way they did, and how that knowledge directly helps you in their interviews.

Amazon, Coupang, and Walmart are three of the largest e-commerce platforms in the world. They also run some of the most sophisticated applied AI systems outside of dedicated AI labs — because at their scale, even a 0.5% improvement in a recommendation algorithm translates to hundreds of millions of dollars in revenue.

---

Amazon — The AI Infrastructure Underneath Everything

Amazon doesn't just use AI products — they build the infrastructure that makes AI scale. Understanding their stack isn't just useful for Amazon interviews; it's useful for understanding production AI at scale.

Rufus — Amazon's Conversational Shopping Assistant

Launched in early 2024, Rufus is Amazon's in-app AI assistant that answers shopping questions conversationally. Ask it "what's a good birthday gift for a 10-year-old who likes LEGO?" and it suggests specific products with context.

What's interesting technically: Rufus isn't just an LLM with product search. It's a RAG system layered over Amazon's entire product catalog, customer reviews, and product Q&As. The retrieval problem is non-trivial — "good birthday gift for a 10-year-old" is a semantic query over 300 million products. The chunking and embedding strategy for product data (title, description, reviews, specifications) is a real engineering problem.

For interviews: if you join an Amazon AI team working on customer-facing products, you'll encounter these problems. "How would you improve Rufus's retrieval precision for vague queries?" is the kind of question that shows up in Amazon AI interviews.

The Recommendation Engine — Item2Vec and Two-Tower

Amazon's "Customers who bought X also bought Y" and "Recommended for you" are powered by multiple models that have evolved significantly since Amazon popularized collaborative filtering.

The current architecture (as described in Amazon research papers):

• Item embeddings via Item2Vec (word2vec applied to purchase sequences — if users often buy A, then B, then C, A and B are "close" in the embedding space)
• User-item interaction modeling via a two-tower neural network (separate towers for user features and item features, trained to maximize dot-product similarity for purchased items)
• Real-time re-ranking using contextual features (current session behavior, time of day, inventory levels)

The scale challenge: 300 million products × 300 million customers means you can't compute every user-item score. The solution: approximate nearest neighbor search (Faiss on item embeddings) for candidate generation, then a lighter model for final ranking.

Supply Chain Optimization — Where OR Meets ML

Amazon's warehouse and delivery optimization is where operations research (OR) and machine learning intersect. Demand forecasting uses gradient boosted trees trained on 3+ years of purchase history, seasonality, and external signals (weather, events). The output of demand forecasting feeds into inventory placement decisions — which products to stock at which fulfillment centers to minimize delivery time and cost.

The real-time routing problem (which of 400,000 delivery routes is optimal right now?) uses a combination of classical OR solvers and ML models trained on historical traffic patterns. This hybrid approach is what makes Amazon delivery windows accurate.

AWS AI Services — Building vs Buying

Amazon's AI infrastructure work shows up publicly through AWS: SageMaker (ML platform), Bedrock (foundational model API), Trainium/Inferentia (custom AI chips). If you're interviewing for an AWS AI team, you're working on the infrastructure millions of companies use to run their own AI systems.

---

Coupang — Speed as the Product, AI as the Engine

Coupang's core product promise is Rocket Delivery: order by midnight, receive by 7am. That's not a logistics promise — it's an AI promise. The system that makes overnight delivery reliably possible is one of the most sophisticated demand forecasting and logistics optimization systems in e-commerce.

Demand Forecasting at SKU × Location Level

Coupang doesn't forecast demand at the product category level — they forecast at the individual SKU × fulfillment center level, for every product, updated multiple times per day. A tube of Crest toothpaste in the Seoul Songpa fulfillment center has its own demand forecast separate from the same product in Busan.

The model architecture (described in engineering posts from the Coupang tech blog): gradient boosted trees for the baseline forecast (captures seasonality, promotions, historical patterns) combined with neural sequence models for capturing recent trend signals. The features include: 30/60/90-day moving averages, day-of-week and holiday effects, competitor promotions, and real-time search volume signals.

The operational challenge: forecasting errors cause two expensive outcomes — stockouts (lose the sale, damage customer trust) or overstock (holds working capital, needs markdown). The loss function isn't symmetric. They optimize for asymmetric loss because the cost of a stockout > cost of overstock in their model.

Real-Time Fraud Detection

Coupang processes millions of transactions daily across South Korea. Their fraud detection runs a two-stage system: a fast classifier (gradient boosted tree, under 10ms) for synchronous transaction approval, and a deeper neural network for asynchronous analysis of suspicious patterns.

The feature engineering challenge: many fraud signals are relational (this card was used by 5 accounts in 24 hours, this account shipped to 3 different addresses in a week). This requires graph features — at what point does a pattern become suspicious? These graph-based features are computed in a near-real-time pipeline and materialized in a feature store.

Product Quality Ranking and Catalog Management

When a customer searches for "running shoes," Coupang has to rank thousands of results. Beyond collaborative filtering-based recommendations, they run a product quality scoring model that considers: return rate, review sentiment (Korean NLP), seller reliability score, product completeness (are all images high quality? is the description complete?). The resulting quality score is a feature in the search ranking model.

For AI interviews at Coupang: they're particularly interested in candidates who understand real-time ML serving (their delivery promise requires sub-second decisions), and who have thought about the operational complexity of running ML at the scale of their logistics network.

---

Walmart — Reinventing a Legacy Giant with AI

Walmart has 10,500 stores and 1.6 million US employees. Their AI transformation story is different from Amazon or Coupang — they're adapting a 60-year-old physical retail operation, not building AI-first from the ground up. That creates a different set of engineering challenges.

Inventory Optimization — The Physical World Problem

Walmart's stores carry 142,000 unique SKUs each. Predicting demand at the store × SKU level — accounting for local demographics, weather, local events, competitor pricing, seasonal patterns — is a massive forecasting problem.

The ML system Walmart built (described in their ICML and KDD papers): a hierarchical forecasting model that shares patterns across similar stores and products (a "global" model with location-specific fine-tuning). This hierarchy is important because smaller stores have less historical data — borrowing statistical strength from similar stores improves forecast accuracy for low-volume products.

The operational feedback loop is physical: when the model predicts high demand for bottled water in a specific store before a hurricane, it needs to trigger not just a purchase order but a supply chain action (route additional trucks, alert the store manager). The ML output connects to an operational action system.

GenAI for Associates — "My Assistant"

In 2024, Walmart launched "My Assistant," a GenAI tool for their 1.6 million US employees. Store associates can ask questions about HR policies, product information, store procedures. Corporate employees use it for document drafting and analysis.

The engineering challenge: a RAG system serving 1.6 million users in real-time, where the knowledge base includes HR policy documents, product manuals, store operations guides, and compliance materials — all of which update frequently. Index freshness, multi-language support (Spanish is critical for Walmart's workforce), and hallucination prevention on policy questions (wrong answer about HR benefits = legal liability) are the hard problems.

Grocery Demand Forecasting — Spoilage as a Loss Function

Grocery adds a dimension that non-perishable retail doesn't have: spoilage. Overstock a product like fresh strawberries and it becomes waste. The demand forecasting loss function must account for spoilage cost, which varies by product (berries spoil faster than apples).

Walmart's grocery demand models incorporate: shelf life of each product category, markdown strategies (reduce price before expiry to reduce waste), and real-time sales velocity signals that trigger early markdowns when a product is selling slower than forecast.

---

What This Means for Your Interviews

Knowing these systems changes how you answer questions at these companies:

At Amazon: When asked "design a recommendation system," don't just describe collaborative filtering. Mention item embeddings, two-tower models, candidate generation via ANN, and real-time re-ranking. Show you understand the scale constraints — you can't compute all user-item scores.

At Coupang: When asked about ML serving, emphasize latency. Coupang's Rocket Delivery promise makes every millisecond count. Show you understand the feature store pattern (fast online store for serving, slow offline store for training) and why it matters for their use case.

At Walmart: When asked about data challenges, acknowledge the physical world complexity. Their data is messier than pure e-commerce data — manual inventory counts, scanner errors, shrinkage. Show you understand how to handle noisy labels and sparse data for low-velocity SKUs.

Deep-dive: System Design

Company-specific system design: recommendation systems, real-time fraud detection, demand forecasting, and supply chain ML

Study →