Online Inventory Management Systems

Machine learning applications rarely stay static—they evolve. What begins as a simple baseline often grows into a multi-stage system shaped by scale, data complexity, and real-world constraints. In this tech blog, the engineering team at Shopify explains how their product classification system evolved as the platform scaled. The journey unfolds across three distinct stages, each with its own technical character. - Stage one focused on a traditional machine learning baseline: logistic regression with TF-IDF features built purely on product text. It was simple, interpretable, and efficient—a practical starting point. - Stage two introduced a multimodal approach, combining both text and image signals within a single model. This significantly improved accuracy, especially when product descriptions were incomplete or ambiguous. However, it remained largely a task-specific classifier trained on a fixed taxonomy. - Stage three marked a shift toward vision-language models. Instead of simply mapping inputs to predefined labels, these models learn richer semantic representations by aligning images and text in a shared embedding space. This enables deeper product understanding and better generalization as taxonomies evolve and new product types emerge. The key takeaway is that real-world machine learning systems mature in layers. You don’t jump straight to the most sophisticated model. Instead, you iterate—balancing accuracy with scalability—and design systems that can adapt as the business grows. #DataScience #MachineLearning #Classification #Evolution #Iteration #SnacksWeeklyonDataScience – – –  Check out the "Snacks Weekly on Data Science" podcast and subscribe, where I explain in more detail the concepts discussed in this and future posts:    -- Spotify: https://lnkd.in/gKgaMvbh   -- Apple Podcast: https://lnkd.in/gFYvfB8V    -- Youtube: https://lnkd.in/gcwPeBmR https://lnkd.in/gYuU_dNT

Types of products and price points and why they exist in a brand's world: Entry Product: Gets someone into the ecosystem. Low risk, high trial. If someone buys the entry product and loves it, you have converted a skeptic into a customer. From there, everything else in the line becomes more accessible. Core Product: The thing most people actually buy. This is where the real business lives. Repeat purchase, real margin, the thing your operations are built around. The core product benefits directly from the entry product pulling people in and the aspirational products anchoring the price. Aspirational Product: Sells very little but does a lot to sell other things. This product exists to define what the brand is capable of and to reframe everything below it. When a brand has a $1,200 product, their $300 product feels like attainable. Without that anchor, the $300 product is just an expensive thing. Halo Product: May not even be (really) for sale. The halo product exists at the ceiling of what the brand will ever do. It is the concept car, the one-of-one, the collaboration that is not for general release. Its job is to define the ceiling and make everything below it feel real and achievable. District Vision dropping a $10k bike, Vollebaks several thousand dollar metal jackets, Satisfy sold a car from a recent brand shoot, the Loro Piana Gift of Kings Wool. The data in the video is from Particl , more info on them in my profile.

Fantastic overview from Shopify explaining how they use vision-language models for product classification. Shopify's product classification system serves as the basis for search ranking and related-product recommendations, so identifying nuance among related products is important. Shopify states that its first system used logistic regression and TF-IDF (a bag-of-words technique that scores words for importance based on frequency across a corpus and a document). This approach lacked classification depth since it was unimodal: it only used a product's description (text) as an input. Shopify determined that its classification system needed to capture more granular product details while adhering to its internal taxonomy, which spans 10,000 product categories. Using a large multimodal model for product classification enables the learning of more intricate relationships across modalities (image + text simultaneously) while also providing zero-shot classification capabilities. To build this system, Shopify implemented FP8 quantization (floating point precision reduction) to reduce memory footprint during inference, on-demand batching with NVIDIA Dynamo to classify products as they surface (versus waiting for an entire batch to fill), and key-value caching. The pipeline itself runs a two-stage prediction: one model call to yield the product's category and another (dependent on the first) to yield its attributes. The pipeline has utilized different models over time; the blog post indicates that the system currently uses Qwen2VL 7B. The pipeline processes 30MM predictions daily and has achieved an 85% classification acceptance rate from merchants. Shopify also built a system to produce training data that sources inputs from multiple LLMs, along with a custom-built model for tie-breaking. Blog post linked below.

These are "Pencil sets" but are classified under different HS codes. In the case of Item ① The description interprets the wooden box as merely serving the purpose of grouping, protecting, and displaying the pencils. However, the pencils constitute the primary component of the product, while the wooden box plays only a supplementary role. Therefore, the writing instruments are considered to be the essential character of the entire product. In the case of Item ② This item also consists of a pencil and its case. However, unlike Item ①, the case is more durable and sturdy.  It is designed to store pencils and accommodate other items, making it a robust and versatile storage solution for various purposes. In such cases, the case can be considered the essential character of the product rather than the pencil itself. Generally, simple packaging used for retail sets is treated as part of the product under GRI 5, and the packaging itself is not classified under an HS code. However, it is important to be aware that there are exceptions, such as Item ②. ①DEBTI37664/24-1(9609.10) https://lnkd.in/g9jvqKtq ➁CZBTI51/271671/2024-580000-04/01(4202.92) https://lnkd.in/gE32TPxn The method for Visualised research HS classification cases is available here: https://lnkd.in/gjw5sXs3 Here is the method of image search for HS classification. Master HS Code Classification Through Images—My Go-To Technique for Daily Article Creation. Image source: © European Union, CC BY 4.0,

🎯 Inventory Classification Systems: Are You Organizing Smart or Just Stocking Blindly? 📦 Have you ever wondered if your inventory is actually working for you—or against you? You may have perfect stock counts… but what about stock value, movement, criticality, or supply risk? Let us spark a quick brainstorm 🔄 How do YOU decide what to prioritize in your warehouse or supply chain? Spoiler: One method is not enough. That is why smart businesses apply multiple classification systems depending on purpose. Here is a breakdown of the most commonly used Inventory Classification Systems 👇 1️⃣ ABC Analysis – Based on consumption value ↳ A: High-value items, tight control ↳ B: Moderate-value items ↳ C: Low-value items, bulk managed 2️⃣ XYZ Analysis – Based on demand variability ↳ X: Predictable demand ↳ Y: Fluctuating demand ↳ Z: Irregular demand 3️⃣ HML Analysis – Based on unit price ↳ H: High-cost items ↳ M: Medium-cost items ↳ L: Low-cost items 4️⃣ FSN Analysis – Based on consumption frequency ↳ F: Fast-moving ↳ S: Slow-moving ↳ N: Non-moving 5️⃣ VEIN Analysis – Based on inventory function ↳ V: Vital ↳ E: Essential ↳ I: Important ↳ N: Not needed / Obsolete 6️⃣ VED Analysis – Based on criticality in operations ↳ V: Vital ↳ E: Essential ↳ D: Desirable 7️⃣ GOLF Analysis – Based on procurement source ↳ G: Government supply ↳ O: Open market ↳ L: Local purchase ↳ F: Foreign source 8️⃣ SDE Analysis – Based on availability or scarcity ↳ S: Scarce ↳ D: Difficult to obtain ↳ E: Easily available 9️⃣ SOS Analysis – Based on seasonality of demand ↳ S: Seasonal ↳ O: Off-seasonal ↳ S: Steady demand throughout the year 🔍 Each method offers a different lens—cost, risk, speed, source, importance, and more. Combining them leads to a more balanced and agile inventory strategy. 💬 What combination of these do you use in your operations? Are there others you would add to the list? Let us share and learn in the comments 💡👇 #Operations #SupplyChain #Logistics #InventoryManagement #WarehouseManagement #Lean #InventoryAnalysis

Understanding ABC Categorization for Effective Inventory Management ABC categorization is a method used to classify inventory into three distinct categories based on their value, usage frequency, and overall importance to the business. This approach helps businesses prioritize inventory management efforts, optimize resources, and improve operational efficiency. Categories of ABC Inventory: A-Category (High Value, Low Volume): These are critical, high-value items that may have a low demand frequency. While they represent a significant portion of the total inventory value, they are often stocked in smaller quantities. Examples include specialty chemicals, high-cost machinery, or unique equipment parts. B-Category (Medium Value, Medium Volume): Items in this category have a moderate value and experience regular demand. They are crucial for everyday operations but do not require the same level of focus as A-category items. Examples include standard machinery parts or common raw materials. C-Category (Low Value, High Volume): These items are typically low in value but are in high demand and consumed in large quantities. They are essential for production and operations but don’t require the same attention as higher-value goods. Examples include fasteners, screws, and packaging materials. Benefits of ABC Categorization: Improved Inventory Management: By classifying inventory into categories based on value and importance, businesses can focus resources and management efforts on high-value, low-volume items that require more frequent monitoring. Reduced Inventory Costs: ABC categorization helps minimize excess stock and reduces waste, obsolescence, and carrying costs, especially for low-value, high-volume items that don’t need as much attention. Enhanced Supply Chain Efficiency: This system streamlines procurement, production, and distribution processes by enabling businesses to prioritize purchasing and stocking strategies based on category importance. Implementing ABC Categorization: Analyze Inventory Data: Review inventory data to understand usage patterns, item values, and demand frequencies. This data forms the foundation for categorizing inventory. Categorize Items: Based on the analysis, assign items into A, B, or C categories according to their value, frequency of use, and importance to the business. Adjust Inventory Levels: Based on the categorization, adjust stock levels, reorder points, and stocking strategies. High-priority A-items should be stocked more carefully, while C-items can be ordered in bulk to meet high demand.

"How do I know which category I belong in?" It sounds simple, but getting this wrong can tank your retail success before you even start. Here's my framework for product categorization: 1️⃣ Start with your product descriptor: This isn't your brand name or flavor. It's the 2-3 words that tell someone exactly what your product is. The key? Your descriptor should NATURALLY categorize your product (+ it needs to be on your packaging right on the front): - "Plant-based yogurt" → dairy/yogurt section - "Collagen powder" → supplements - "Date sugar" → baking/sweeteners 2️⃣ Understand your customer use case: This is CRITICAL and often overlooked. Where would your customer logically look for your product or look to use your product? Example: A protein cookie could go in: → Supplements (for fitness folks) → Natural snacks (for healthy snackers) → Cookie aisle (for better-for-you treat seekers) The right answer? Wherever your core customer would look first! 3️⃣ Define your value-add: Who are you an alternative for and why? This helps pinpoint your exact category placement. For example: - Are you a healthier alternative? → Natural section - A premium option? → Specialty section - A convenient version? → Grab-and-go area 4️⃣ Validate with retail visits: Double-check your choice by doing the research.→ Visit 3+ retailers (mix of conventional & natural) → Check Amazon & Walmart's categorization → Map how each organizes similar products 🌟 Reality check: 🌟 Your category isn't just about where YOU think you belong. It's about: - Where buyers will place you - Where customers will look for you - Where you'll actually succeed on shelf Pro tip: Your category might shift slightly between retailers. That's ok! What matters is that you understand how each retailer organizes their products, so you can speak their language and meet your customer where they shop.

A framework I use to predict categories before they come about. Hopefully it helps you spot the next ones. Great brands is born from one of two instincts: - To disrupt (category disruption) - To create (category creation) Category Disruption When you disrupt, you’re not changing human desire, just the path to fulfilling it. And there are patterns to this game. - WYSIWYG brands: where the sum of parts equals truth. People know they’re paying a “brand premium” because parts cost far less than the price. OnePlus broke open the black box: “Premium is just pricing illusion.” - Access/Process Simplifiers: when people wanted the product but didn’t know how or where to buy it. → Access: Our parents shopped in markets; we shop in malls. They knew which store sold what; we don’t. P-Tal and House of Chikankari filled that gap. → Process: Lenskart made buying glasses simple. - Trust Simplifiers: when buyers didn’t trust the market. Standardization was missing; players were grey. And Nothing Else and P-Tal didn’t invent brass utensils, they standardized trust. - New-Age Reclaimers: where old brands lost emotional resonance. “I don’t want to use the brand my dad did.” Mokobara, Bombay Shaving Co., Pant Project. - Underserved Segment Builders- where incumbents ignored a segment. boAt saw one large player and built for the premium-aspirational middle. The more truths you hit (brands can hit more than one), the deeper your wedge into culture. Once you spot the gap, you have three levers: Premiumize. Undercut. Or Go Mass. -Undercut: When incumbents have bloated margins- offer a premium experience at a mass price. boAt, OnePlus. Mostly WYSIWYG categories. Premiumize: When a category feels stale- elevate product, trust, or access. → Experience: In completely commoditized categories. Chaayos premiumized tea. → Product: Standardize (P-Tal), add trust (And Nothing Else), or help people differentiate (Mokobara), and you can charge a premium. Go Mass (Consolidate): When the market’s fragmented with no clear winner. You don’t premiumize or undercut- you simplify, standardize, and scale. Think Lenskart. Price isn’t the strategy. It’s the story your audience tells themselves about value. Category Creation Creating a new category is different. You’re not disrupting behavior - you’re teaching it. That’s why you start premium. Mass markets don’t buy innovation; they buy proof it works. So you premiumize first, democratize later. In new categories0 Fitbit, Tesla- you start premium, then go mass. Early adopters have the problem and will pay for you to solve it. Category creation usually follows a tech or policy shift. Sometimes tech even creates price advantages, lab-grown diamonds, for instance. But even then, brands target the premium segment first, those who understand and believe the shift. Even if it’s cheaper than diamonds, it’s sold to educated, well-earning India. Some brands re-imagine the old, some invent the new. All great ones seem inevitable.

One thing most recommender systems lack is a clean way to represent what an item actually is. Most production systems rely on item IDs like 12345, while LLM-based approaches often generate free-form descriptions that are inconsistent. The same product might be tagged as Blues Music in one place and Classic Blues Hits somewhere else. That makes the representations difficult to reuse and hard to control. A recent paper, AgenticTagger from Google, proposes a different approach. Instead of asking an LLM to describe every item independently, it uses a team of LLM agents to build a shared taxonomy of tags and then annotates every item using only those approved tags. The end result is a structured hierarchy such as Rock → Glam Rock → Proto Punk, where every item follows the same vocabulary. The framework consists of two types of agents. The first is an Architect. Its job is to create the taxonomy by defining categories and subcategories. The second is a group of Annotators. They receive batches of products and assign them to the existing taxonomy. When they encounter an item that does not fit anywhere, they generate feedback for the Architect instead of inventing a new label. For example, suppose an Annotator encounters a 0-degree sleeping bag, but the taxonomy only contains a generic Sleeping Bags category. Rather than creating an arbitrary label, the Annotator reports the gap. The Architect then updates the taxonomy with a new category such as Cold Weather Sleeping Bags. Over multiple iterations, the taxonomy expands only where needed, resulting in a vocabulary that is both comprehensive and consistent. The authors evaluated this iterative process across Amazon datasets covering Sports, Beauty, and CDs. In the Beauty category, for example, the system constructs a hierarchy like Haircare → Flat Irons & Straighteners → Ceramic Plate Irons. These structured tags are then used as features for downstream recommendation models. Compared to traditional item IDs and free-form LLM descriptions, the structured taxonomy consistently improves recommendation quality. Since similar products now share meaningful tags, the model can better generalize across related items, even when a product has very little interaction history. IMO, the practical value of this paper is not just better recommendation accuracy. It gives recommender systems a vocabulary that is both interpretable and controllable. Users can express preferences like show me more Jazz, and the system can directly retrieve items associated with the Jazz tag instead of relying on opaque embeddings or inconsistent descriptions. From an engineering perspective, the approach is also easy to integrate into existing production systems. These hierarchical tags behave like standard categorical features, allowing teams to leverage LLM-generated semantic understanding without replacing their existing recommendation infrastructure. Link to Paper: https://lnkd.in/gYMwQi7J

Are you treating all inventory the same? If so, you're tying up cash and hurting service! ABC analysis classifies products by value. XYZ analysis shows how predictable demand is. Combined, they are a powerful tool to examine your product portfolio. ABC analysis: 🟢 A: Highest value items. Typically 80% of sales, but only around 10-20% of SKUs. 🟡 B: Moderate value items: Middle 15% of sales, usually around 20-30% of SKUs. 🔴 C: Low value items: Bottom 5% of sales, usually 50-70% of SKUs. XYZ analysis: 🔵 X: Stable, predictable demand. High forecast accuracy. Core products. 🟣 Y: Moderate demand variability. Some seasonality or trends. Seasonal or promotional items. ⚫ Z: Highly variable or sporadic demand. Difficult to forecast. One-off orders. Together, they guide decisions like: ✅ Setting service level targets of 99.9% for AX items, decreasing to 92% for CZ items. You never run out of your top sellers, hit 95% overall service, and still dramatically slash inventory. ✅ Deciding to make BZ, CY, and CZ items to order instead of keeping them in stock. You free up a ton of cash without much impact on sales. Or work with customers get orders in earlier, or price in carrying costs if they can't. ✅ Eliminating SKUs that look profitable on paper but distract from your core business. Inventory segmentation isn't just about service levels. It's a growth strategy. Free up cash, focus on what sells, and fuel your business forward.