<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Arav Agnihotri on Medium]]></title>
        <description><![CDATA[Stories by Arav Agnihotri on Medium]]></description>
        <link>https://medium.com/@aravagnihotriofficial?source=rss-55766841452d------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*PoxGKJNJDHV6fg_cvy7WLA.jpeg</url>
            <title>Stories by Arav Agnihotri on Medium</title>
            <link>https://medium.com/@aravagnihotriofficial?source=rss-55766841452d------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 06:15:22 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@aravagnihotriofficial/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Shopping is broken. Here’s how I am aiming to fix it]]></title>
            <link>https://medium.com/@aravagnihotriofficial/shopping-is-broken-heres-how-i-am-aiming-to-fix-it-88201cc96e4c?source=rss-55766841452d------2</link>
            <guid isPermaLink="false">https://medium.com/p/88201cc96e4c</guid>
            <dc:creator><![CDATA[Arav Agnihotri]]></dc:creator>
            <pubDate>Thu, 28 May 2026 23:33:49 GMT</pubDate>
            <atom:updated>2026-05-28T23:33:49.138Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>By Arav Agnihotri · The Knowledge Society · TKS Focus Project, 2026</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_sLlrEQt-eCf0Se7wIon4g.png" /></figure><p>Every time you open a shopping site, you’re handed a search bar and a prayer. You type in something like “casual summer dress” and get back 4,000 results, most of which are vaguely related to what you wanted and none of which account for the fact that you’re 5&#39;4&quot;, have a specific idea of what “casual” means to you, and just bought three things in the same color last month.</p><p>The filters don’t help. Sorting by “relevance” is a black box. And size charts — those infamous size charts — are written by a different team at every brand, measured in completely different ways, and somehow still wrong half the time.</p><p>Returns are the real cost of all of this. In 2023, online shoppers in North America returned nearly $248 billion in merchandise. The primary reason? Products didn’t fit or match expectations. That number isn’t a retail statistic — it’s proof that the entire discovery experience is failing people at the moment that matters most: deciding whether to buy.</p><p>The problem isn’t that people don’t know what they want. It’s that no tool has ever actually listened.</p><p>Current “AI” shopping tools are mostly keyword wrappers. They take your query, punch it into an existing search engine with a different label on it, and hand you the same results you would’ve gotten yourself. They don’t know you. They don’t know your body. They have no concept of what’s already in your closet or what a “good fit” actually means for your proportions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4-uecWY1Y6CIXHe5zNvNpg.png" /></figure><h3>Drayp: a shopping agent that actually knows you</h3><p>Drayp is an AI-powered shopping agent built around a simple idea: what if buying clothes online felt less like browsing a warehouse and more like talking to a stylist who’d known you for years?</p><p>The experience starts with a profile — not a form you fill out and forget, but a living context layer. You upload a photo (validated by GPT-4 Vision to confirm it’s actually a usable image of you), enter your body measurements, and set your style preferences: the brands you reach for first, the colors you keep coming back to, the aesthetic tags that actually describe how you dress.</p><p>From there, the main interface is a chat. You talk to Drayp the same way you’d text a friend: “I need something for a rooftop dinner next Saturday, kind of dressed up but not stiff” or “find me joggers that would work with the sneakers I already have.” The agent parses that into structured intent — category, occasion, vibe, formality, constraints — and starts working.</p><p>What it does next is where it gets different. Instead of keyword-matching your query against product titles, Drayp uses semantic search: your intent is embedded into a vector and compared against embedded product descriptions using cosine similarity. Products that mean the same thing as what you asked for surface, even if they use completely different words to describe it.</p><p>Every result comes with a fit score — a 0 to 100 prediction of how well that specific item will actually fit your body, calculated from your measurements against the product’s sizing data. Not “this is a medium.” A real number, grounded in your actual dimensions.</p><p>And then there’s the Try On feature. For any product you’re considering, Drayp generates an AI image of you wearing it — using your uploaded photo as a reference — and feeds that output into a 3D modeling pipeline via Trellis to produce a rotatable 3D render. You see yourself in the clothes before you buy them. Not a model who shares your rough vibe. You.</p><h3>How it’s built</h3><p>Drayp runs on Next.js 14 with the App Router, written in TypeScript, and styled with Tailwind CSS. Authentication is handled by Clerk; the database is Supabase. The whole stack is built to be deployed fast and iterated on faster.</p><p>The agent logic is the core of everything. When a user sends a message, the system converts it into structured intent — extracting the product category, occasion, style tags, and any hard constraints from natural language. That intent object then drives a product search via Serper.dev, which I chose specifically because it delivers significantly more search volume per dollar than alternatives like SerpAPI. At the scale I need for testing and iteration, that margin matters.</p><p>Once products are retrieved, they don’t just get handed back raw. Each result is passed through a semantic re-ranking layer: descriptions are embedded using OpenAI’s text-embedding-3-small model, the user&#39;s intent is embedded the same way, and cosine similarity scores determine the final ranking. The products that genuinely match what you asked for — not just lexically, but semantically — rise to the top.</p><p>The fit score algorithm pulls from the user’s saved body measurements and maps them against the sizing data available in each product listing. The output is a single number per product. The goal is to make it feel instinctive: an 87 means this thing will probably feel great on you; a 43 means you’d be taking a risk.</p><p>The Try On pipeline is the most technically interesting piece. When a user requests a try-on, their profile photo and the product image are sent to a generation model that composites the two — creating a realistic image of that specific person in that specific item. That generated image is then passed to Trellis, hosted on Replicate (since Trellis doesn’t expose a standalone callable API), which produces a 3D model the user can rotate in-browser.</p><p><strong>The full tech stack:</strong></p><ul><li>Next.js 14 — framework and routing</li><li>TypeScript — language</li><li>Tailwind CSS — styling</li><li>Clerk — authentication</li><li>Supabase — database</li><li>GPT-4 Vision — photo validation</li><li>OpenAI text-embedding-3-small — semantic search and re-ranking</li><li>Serper.dev — product search</li><li>Replicate + Trellis — 3D try-on pipeline</li></ul><h3>What’s next</h3><p>The initial monetization path is affiliate commissions — standard for a shopping product, but also a quick way to validate that Drayp is actually driving purchasing decisions, not just browsing sessions. If people buy through it, the model works.</p><p>From there, the Try On feature becomes a freemium gate. The core agent is free; unlimited 3D try-ons are a paid tier. That’s the bet: the feature with the highest production cost is also the one that makes the biggest difference in purchase confidence. It should be worth paying for.</p><p>Beyond that, the long play is brand partnerships. A brand that knows Drayp users are high-intent, measurement-verified, style-profiled shoppers has a very different kind of advertising surface than a typical display unit. That’s the moat.</p><p>Right now, I’m completing the full initial build in Cursor — all pages, API routes, database schema, and environment wiring — and iterating on the fit score algorithm and try-on pipeline quality. The goal is a working product I can put in front of real users and measure.</p><p>The problem of online shopping feeling impersonal and inaccurate is big, it’s measurable, and it costs real money every year. Drayp is my attempt to fix it — one conversation at a time.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=88201cc96e4c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How My AI Weather Email Agent Works]]></title>
            <link>https://medium.com/@aravagnihotriofficial/how-my-ai-weather-email-agent-works-55dd18c88ac4?source=rss-55766841452d------2</link>
            <guid isPermaLink="false">https://medium.com/p/55dd18c88ac4</guid>
            <dc:creator><![CDATA[Arav Agnihotri]]></dc:creator>
            <pubDate>Thu, 05 Feb 2026 17:25:17 GMT</pubDate>
            <atom:updated>2026-02-05T17:25:17.896Z</atom:updated>
            <content:encoded><![CDATA[<h3>How My AI Weather Email Agent Works (Built with n8n)</h3><p>I built an automated AI agent that sends a daily weather report straight to my inbox — without me doing anything. It combines scheduling, live weather data, memory, and AI reasoning into one seamless workflow.</p><p>Here’s exactly how the system works from start to finish.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1C84u-eogmprMHafhIT8xg.png" /></figure><h3>1. Schedule Trigger — The System’s Alarm Clock</h3><p>Every automation needs a starting point.<br>The <strong>Schedule Trigger</strong> is what wakes the workflow up.</p><p>At a set time each day, this trigger activates the entire system. No manual input is required — the process runs automatically in the background.</p><h3>2. AI Agent — The Brain of the Operation</h3><p>The <strong>AI Agent</strong> is the decision-maker and coordinator. Instead of hard-coding steps, I gave it tools and let it figure out how to use them.</p><p>It has access to:</p><ul><li>A language model (OpenAI)</li><li>Memory of previous runs</li><li>External tools (weather + calendar)</li></ul><p>When triggered, the agent’s goal is simple:<br><strong>“Gather today’s weather and generate a well-formatted email report.”</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fMs60VpGO_YkUuOiwObHeg.png" /></figure><h3>3. OpenAI Model — Writing the Email</h3><p>The AI Agent connects to an OpenAI model, which is responsible for:</p><ul><li>Turning raw weather data into natural language</li><li>Formatting the report as clean HTML</li><li>Writing a proper subject line and email body</li></ul><p>This is what transforms structured data into something that feels like a human-written update.</p><h3>4. Weather Tool — Getting Live Conditions</h3><p>The agent uses the <strong>Current Weather</strong> tool to pull real-time weather data.</p><p>This includes information like:</p><ul><li>Temperature</li><li>Feels-like temperature</li><li>Highs and lows</li><li>Wind speed and direction</li><li>Humidity</li><li>General conditions (cloudy, clear, etc.)</li></ul><p>The agent does not “guess” the weather — it retrieves actual live data and then explains it.</p><h3>5. Memory — Keeping Context Between Runs</h3><p>The <strong>Simple Memory</strong> node allows the agent to retain context across executions.</p><p>That means over time the system can:</p><ul><li>Stay consistent in formatting</li><li>Potentially compare day-to-day changes</li><li>Improve continuity in reports</li></ul><p>It turns the agent from a one-time responder into a system that can evolve.</p><h3>6. (Optional) Calendar Tool</h3><p>The calendar tool is available to the agent as well. While not required for basic weather reports, it allows the system to expand later — for example:</p><ul><li>Adding weather notes to calendar events</li><li>Adjusting reports based on scheduled activities</li></ul><p>This makes the agent extensible, not just single-purpose.</p><h3>7. Email Output — Delivering the Result</h3><p>After generating the report, the AI Agent outputs structured data containing:</p><ul><li>A <strong>subject line</strong></li><li>A <strong>fully formatted HTML message</strong></li></ul><p>That output is passed into the <strong>Gmail Send Email</strong> node.</p><p>Because the data is structured correctly, Gmail can directly use:</p><ul><li>subject → email subject</li><li>message → email body (rendered as HTML)</li></ul><p>And just like that, the system sends a polished daily weather email automatically.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5aaVoSWNNWOwgZ1grMAtjA.png" /></figure><h3>What Makes This Powerful</h3><p>This isn’t just a script. It’s an <strong>AI-driven system</strong> that:</p><ul><li>Decides which tools to use</li><li>Pulls live external data</li><li>Writes human-style communication</li><li>Executes real-world actions (sending emails)</li></ul><p>All on a schedule, with no human in the loop.</p><p>It’s a small example of something much bigger:<br><strong>AI agents that observe the world, think, and take action.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=55dd18c88ac4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Replicating SMPL: Understanding the Foundations of Digital Human Models]]></title>
            <link>https://medium.com/@aravagnihotriofficial/replicating-smpl-understanding-the-foundations-of-digital-human-models-8b104e609e2b?source=rss-55766841452d------2</link>
            <guid isPermaLink="false">https://medium.com/p/8b104e609e2b</guid>
            <dc:creator><![CDATA[Arav Agnihotri]]></dc:creator>
            <pubDate>Sun, 02 Nov 2025 22:58:17 GMT</pubDate>
            <atom:updated>2025-11-02T23:47:05.523Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/480/1*7S5qxPiyDgQxegnU_XI1Ig.png" /><figcaption>The SMPL parametrization encodes 3D body shape, pose, soft tissue motion, hands, fingers, facial expressions — <a href="http://meshcapade.me/">http://meshcapade.me/</a></figcaption></figure><h3>Introduction</h3><p>Agentic Artificial intelligence increasingly relies on accurate digital representations of the human body for applications in retail, robotics, and healthcare. The Skinned Multi-Person Linear (SMPL) model provides a mathematical framework for generating realistic 3D human forms from a compact set of numerical parameters.</p><p>This project replicates a simplified part of Meshcapade’s SMPL system to understand how human body shape can be encoded, modified, and rendered using open-source tools. The goal wasn’t to recreate their full pipeline, but to break down its structure and show how mathematical coefficients can be turned into a realistic 3D human mesh.</p><p>Check my project out here: <a href="https://3d-model-arav-agnihotri.replit.app/">https://3d-model-arav-agnihotri.replit.app/</a></p><h3>Project Overview</h3><p>The implementation converts numerical parameters describing body structure — specifically <em>shape coefficients</em> (β) and <em>joint rotations</em> (θ) — into a three-dimensional mesh. The resulting mannequin represents the human form without textures or colors. Users can view the mesh in solid or wireframe mode, rotate it, and examine the effects of changing parameter values.</p><p>This foundational process underlies many technologies, including virtual try-on systems, ergonomic analysis, and biomechanics simulation.</p><h3>The process:</h3><h4>Environment Setup</h4><p>The project was developed on <strong>Replit</strong>, an online IDE that supports real-time execution and visualization. The following Python libraries were installed:</p><pre>!pip install numpy matplotlib open3d pickle5</pre><p><strong>Open3D</strong> was used for 3D visualization because it allows interactive inspection of meshes directly within the browser.</p><h4>Loading the SMPL Model</h4><p>Official SMPL .pkl files were downloaded; these contain the model’s vertex templates, blend shapes, and joint regressors. The files were loaded as follows:</p><pre>import pickle5 as pickle<br>import numpy as np</pre><pre>with open(&quot;SMPL_NEUTRAL.pkl&quot;, &quot;rb&quot;) as f:<br>    smpl_model = pickle.load(f, encoding=&quot;latin1&quot;)</pre><pre>vertices_template = smpl_model[&quot;v_template&quot;]<br>shapedirs = smpl_model[&quot;shapedirs&quot;]<br>posedirs = smpl_model[&quot;posedirs&quot;]<br>J_regressor = smpl_model[&quot;J_regressor&quot;]<br>weights = smpl_model[&quot;weights&quot;]<br>faces = smpl_model[&quot;f&quot;]</pre><p>The <strong>shape directions</strong> (shapedirs) encode how the mesh deforms as body shape parameters change, while <strong>pose directions</strong> (posedirs) define deformations associated with joint rotations.</p><h4>Generating Body Shape</h4><p>A simple function was written to apply shape coefficients (β values) to the template mesh. Each coefficient slightly modifies body proportions:</p><pre>def apply_shape(betas, shapedirs, v_template):<br>    shaped_mesh = v_template + np.tensordot(shapedirs, betas, axes=[[2], [0]])<br>    return shaped_mesh</pre><p>Testing with random β values allowed observation of how each coefficient influenced the final body geometry:</p><pre>betas = np.random.randn(10) * 0.03<br>body_vertices = apply_shape(betas, shapedirs, vertices_template)</pre><p>This demonstrated how SMPL represents the human body purely through numerical transformations rather than direct imaging.</p><h4>Rendering the Mesh</h4><p>Initial visualization attempts used Matplotlib’s 3D plotting tools, but rendering was limited and slow. Switching to <strong>Open3D</strong> provided a more realistic and responsive interface:</p><pre>import open3d as o3d</pre><pre>def visualize_mesh(vertices, faces):<br>    mesh = o3d.geometry.TriangleMesh()<br>    mesh.vertices = o3d.utility.Vector3dVector(vertices)<br>    mesh.triangles = o3d.utility.Vector3iVector(faces)<br>    mesh.compute_vertex_normals()<br>    o3d.visualization.draw_geometries([mesh])</pre><pre>visualize_mesh(body_vertices, faces)</pre><p>This produced an interactive 3D viewer that supported rotation, zooming, and toggling between solid and wireframe visualization.</p><h4>Testing and Analysis</h4><p>The mesh was regenerated repeatedly with varying β parameters to analyze the effect of each coefficient. For example, increasing certain parameters broadened the chest or elongated the limbs, demonstrating how small numerical variations can produce distinct body morphologies.</p><p>Joint data was also tested for potential pose deformation, though full kinematic implementation was beyond the project’s current scope.</p><h3>Results</h3><p>The replicated SMPL component successfully generated a controllable 3D human model from parametric inputs. The system demonstrated real-time rendering and visualization in Open3D, with measurable, interpretable effects from modifying shape parameters.</p><p>While limited to static poses, the project confirmed the mathematical underpinnings of SMPL and illustrated how digital human representations can be built without direct visual input data.</p><h3>Conclusion</h3><p>This replication exercise provided practical insight into how data, geometry, and computation converge in digital human modeling. The process clarified how linear blend-shape models like SMPL encode body structure efficiently and flexibly.</p><p>Assistance from AI tools such as ChatGPT was used to explain mathematical concepts and coding syntax during development, enabling a deeper conceptual understanding of each module rather than relying on automated generation.</p><p>The project highlighted the scalability of SMPL-based techniques for applications in virtual fitting, motion capture, and healthcare analytics, where privacy-preserving digital avatars could serve as intermediaries between users and AI systems.</p><p>Replicating a small part of the SMPL model demonstrated how the human body can be mathematically parameterized and rendered using open-source tools. Beyond the technical outcome, this project reinforced how structured data can become a foundation for digital embodiment — a concept central to the future of AI-driven retail, robotics, and healthcare.</p><p>Check out my website here: <a href="https://aravs-personal-website.lovable.app/">https://aravs-personal-website.lovable.app/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8b104e609e2b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Agentic E-Commerce: How SMPL-Powered AI Agents Will Revolutionize Retail Shopping]]></title>
            <link>https://medium.com/@aravagnihotriofficial/agentic-e-commerce-how-smpl-powered-ai-agents-will-revolutionize-retail-shopping-d53d736ec470?source=rss-55766841452d------2</link>
            <guid isPermaLink="false">https://medium.com/p/d53d736ec470</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ecommerce]]></category>
            <category><![CDATA[fashion]]></category>
            <category><![CDATA[3d-modeling]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Arav Agnihotri]]></dc:creator>
            <pubDate>Sat, 04 Oct 2025 03:40:33 GMT</pubDate>
            <atom:updated>2025-10-11T20:29:59.245Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Turning smartphones into precise tailoring tools, cutting returns and reshaping global retail.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EbRBfqdqmVDC8PZ6sseYbQ.png" /><figcaption>SMPL model showing joints estimated from weighted mesh vertices.<br> <em>Source: Khanh Ha, “SMPL model introduction” (2020). </em><a href="https://khanhha.github.io/posts/SMPL-model-introduction/"><em>Link</em></a></figcaption></figure><p>Agentic e-commerce is set to change retail at its core. Amazon recently introduced “Buy for Me,” a shopping agent that can search, compare, and complete purchases across Amazon and third-party sites. OpenAI has integrated agentic shopping into ChatGPT, allowing users to buy directly from partners like Etsy and soon Shopify, positioning its assistant as both a conversational tool and a transactional one. Together, these moves show that both companies view autonomous shopping agents as a central part of the future of e-commerce.</p><p>With SMPL-powered agents, shopping moves from a passive search and click process to an autonomous goal driven process where technology acts on behalf of the consumer. Instead of navigating product pages or managing multiple checkouts, customers define intent and the agent handles the rest, searching, comparing, deciding, and completing transactions across platforms. These agents can reduce one of retail’s biggest problems, product returns. By analyzing body measurements, past purchases, sizing charts, and brand specific fit data, SMPL agents help shoppers find the correct fit before buying, making returns less likely. The outcome is lower costs for retailers, who spend billions annually on reverse logistics and lost resale value, and reduced environmental harm, with millions of tons of CO₂ emissions and waste generated each year from returned goods. This represents a significant transformation of retail into an efficient and sustainable system where commerce is personalized and proactive.</p><h3>So, what is SMPL?</h3><p>SMPL, short for Skinned Multi-Person Linear model, is a parametric 3D human body model that has become the standard for representing digital humans in computer vision. Instead of storing thousands of separate body scans, SMPL compresses human variation into two sets of parameters: shape, which defines body proportions such as height, weight, or muscularity, and pose, which defines how joints move and rotate in 3D space. With these inputs, SMPL generates a <a href="https://www.google.com/search?q=mesh+definition&amp;sca_esv=f12af29252359844&amp;sxsrf=AE3TifNcwSJ71BJ8mtvI9rDvxXrb7glQNQ%3A1759547492123&amp;ei=ZJDgaMqkB9jA0PEP5fOSsAg&amp;oq=mesh+def&amp;gs_lp=Egxnd3Mtd2l6LXNlcnAiCG1lc2ggZGVmKgIIADILEAAYgAQYkQIYigUyBRAAGIAEMgUQABiABDIFEAAYgAQyChAAGIAEGBQYhwIyChAAGIAEGBQYhwIyBRAAGIAEMgUQABiABDIFEAAYgAQyBRAAGIAESNIRUP8BWO8KcAJ4AZABAJgBmAGgAeoDqgEDNC4xuAEByAEA-AEBmAIHoAKRBMICBxAjGLADGCfCAgoQABiwAxjWBBhHwgINEAAYgAQYsAMYQxiKBcICDRAAGIAEGLEDGEMYigXCAgoQABiABBhDGIoFwgINEAAYgAQYsQMYFBiHAsICBxAAGIAEGArCAgYQABgWGB7CAggQABgWGAoYHpgDAIgGAZAGCpIHAzUuMqAHzBmyBwMzLjK4B4YEwgcFMC4xLjbIBxs&amp;sclient=gws-wiz-serp">mesh</a> of the body that deforms naturally using a technique called linear blend skinning, the same method used in video games and animation to make characters move smoothly. The model was trained on thousands of high-resolution body scans, so it captures both detailed shape variation and natural movement. This makes SMPL useful for applications in film, AR and VR avatars, motion capture, fitness, sports analysis, and fashion technology. Extensions such as SMPL X expand the model to include detailed hand poses and facial expressions, while variants such as STAR allow faster real-time performance. SMPL functions as a digital template of the human body that can be scaled, posed, and adapted for research, design, or commerce.</p><figure><img alt="SMPL model showing joints estimated from weighted surrounding mesh vertices." src="https://cdn-images-1.medium.com/max/584/1*vCw_zr6lTXqHXiN7q1YuYg.png" /><figcaption>SMPL model with joint regression, where each joint position is estimated as a weighted combination of surrounding vertices.<br> <em>Source: Khanh Ha, “SMPL model introduction” (2020). </em><a href="https://khanhha.github.io/posts/SMPL-model-introduction/"><em>Link</em></a></figcaption></figure><h3>How does it work?</h3><p>SMPL works by combining statistics, geometry, and animation into a single model of the human body. At its foundation, it represents a person as a 3D mesh, a set of points connected into triangles that can be reshaped and posed. To capture the diversity of human bodies, SMPL uses shape parameters (β), a small set of numbers that adjust features such as height, weight, muscle tone, and proportions. These parameters were learned by analyzing thousands of 3D scans and compressing them into a compact representation of human variation. Once the body shape is defined, pose parameters (θ) control the rotation of each joint in the skeleton. This allows the mesh to be animated in different positions, with 72 values covering the major joints of the body. To make these movements smooth, SMPL applies Linear Blend Skinning, a technique where each point on the mesh is influenced by nearby joints, ensuring realistic bending and stretching instead of distortions. SMPL also includes blend shapes, which are learned deformations that model the natural effects of motion on muscles and skin, such as shoulders moving when arms are raised. These elements together transform a set of parameters into a 3D human body that can be customized for different uses.</p><p>SMPL is a model that requires input from real-world data. Sensing and capture technologies provide this input. LiDAR, photogrammetry, and computer vision each offer methods for collecting information about the human body, which can then be translated into SMPL’s shape and pose parameters. Depending on the application, one method may be more suitable than another. LiDAR provides centimeter-level accuracy for medical or fashion applications. Photogrammetry captures high-detail scans that SMPL can process for film or gaming. These pipelines connect raw human data to the 3D models that SMPL generates.</p><p><strong>Sensing and Capture Technologies<br></strong> LiDAR (Light Detection and Ranging) is an active sensing technology, that measures distance by emitting pulses of laser light and recording the return time. The time of flight of each pulse is converted into a distance value, and when repeated millions of times per second, it produces a dense 3D point cloud of the environment. Each point has precise coordinates (x, y, z), mapping the surface geometry of objects at real-world scale. Unlike photogrammetry, which estimates depth by comparing 2D images, LiDAR directly measures it, making it more reliable in challenging lighting or low-texture environments.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bmpioSJixq69g1chjEW95w.png" /><figcaption>Using an iPhone’s LiDAR sensor to create a digital floor plan through room scanning.<br> <em>Source: HomeDevices, “Create an Apple Home floor plan with a LiDAR sensor” (2023). </em><a href="https://homedevices.app/create-an-apple-home-floor-plan-with-a-lidar-sensor/"><em>Link</em></a></figcaption></figure><p>Beginning with the iPhone 12 Pro, Apple integrated a VCSEL-based infrared LiDAR module that operates on the time-of-flight principle. The sensor emits bursts of invisible infrared light (~940 nm wavelength) and measures the return time using a photodiode array. This produces a depth map with tens of thousands of points per frame, updating at up to 60 fps, with accuracy typically between 2 mm and 5 mm at close distances (0.5–3 m). The LiDAR unit works with the iPhone’s RGB camera and Apple’s neural engine to combine color information with depth, generating real-time 3D meshes. The iPhone LiDAR is solid-state, which makes it smaller, more power efficient, and optimized for short-range scanning of people and objects.</p><p>Iphone LiDAR lowers barriers by leveraging a tool that many people already own and understand, eliminating the need for costly hardware. This makes adoptions easy for consumers and scalable for retailers, drivering higher engagement. The technology is effective for capturing the geometry of the human body because it provides absolute scale rather than relative shape. For healthcare or custom apparel, centimeter-level accuracy is important. A person can be scanned by moving around them with the iPhone or by having them rotate slowly in place, producing millions of depth points that can be merged into a watertight mesh. Applications such as Polycam, Scaniverse, and Luma AI process this point cloud by triangulating surfaces, applying RGB textures, and generating a 3D model in formats such as OBJ, STL, or GLTF. When aligned with SMPL, the statistical model refines the noisy point cloud into a clean, rigged, and animatable human mesh while retaining real-world dimensions and incorporating SMPL’s shape and pose parameters.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ipog71DMxaU5AKIsBkoPwg.png" /><figcaption>Comparison of LiDAR (laser-based measurement) and photogrammetry (image-based reconstruction).<br> <em>Source: RVS Land Surveyors, “Lidar or Photogrammetry” (2023). </em><a href="https://www.rvslandsurveyors.com/lidar-or-photogrammetry?utm_source=chatgpt.com"><em>Link</em></a></figcaption></figure><p>Photogrammetry differs from LiDAR in that it does not directly measure distance with laser pulses but instead estimates depth from images. The method requires capturing multiple overlapping photographs of an object or person from different angles. By analyzing how features align across these images, photogrammetry algorithms apply geometric triangulation to reconstruct a 3D surface. While LiDAR measures depth through emitted light and return time, photogrammetry infers depth by comparing pixels and calculating their 3D positions.</p><p>Photogrammetry is based on a structure-from-motion (SfM) pipeline. The software first detects keypoints such as edges, textures, or corners that appear across multiple photos. It then estimates the relative positions of the cameras at the time each image was taken. Using these correspondences, the system triangulates the 3D coordinates of points. The result is a point cloud that can be processed into a dense mesh. Unlike LiDAR, which requires a specialized sensor, photogrammetry only requires a camera. This makes it usable with smartphones or DSLR cameras. High-end systems may employ dozens or hundreds of cameras capturing images simultaneously.</p><p>Photogrammetry is effective at capturing visual detail because it uses RGB imagery, allowing reconstructed 3D models to include high-resolution textures. In body scanning, this enables recovery of surface geometry along with features such as skin tone, clothing patterns, and fine details. When integrated with SMPL, the raw photogrammetry mesh can be aligned with the statistical body model, producing a result that combines accurate body articulation with detailed surface textures. This makes photogrammetry particularly valuable in industries such as gaming, film, and fashion, where visual detail is prioritized over sub-centimeter measurement accuracy.</p><h3>How SMPL and Sensing and Capturing Technologies Create 3D Body Models</h3><p>Once SMPL has input data, whether from a LiDAR scan, a photogrammetry model, or photographs, it does not immediately produce a model. It uses those inputs to progressively reshape a digital mannequin until it aligns with the real body.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XWcCTiaN9_RNyQv--h0ubQ.png" /><figcaption>Examples of the SMPL body model illustrating variations in body shape and pose.<br> <em>Source: Max Planck Institute for Intelligent Systems, “SMPL: A Skinned Multi-Person Linear Model” (2015). </em><a href="https://is.mpg.de/ps/code/smpl"><em>Link</em></a></figcaption></figure><p><strong>1. Starting Template<br></strong> SMPL begins with a standard body template in a neutral standing pose with arms slightly apart. This template is represented as a mesh composed of thousands of triangles that form a surface. The fixed template provides consistency across subjects. If a measurement such as height is available, SMPL scales the template so that the model operates in real-world units. All subsequent steps, including pose adjustment, shape refinement, and error correction, build from this foundation.</p><p><strong>2. Pose Matching<br></strong> The next stage is to match the posture of the subject. SMPL represents the human body with 24 joints, each with three rotational degrees of freedom, producing 72 pose parameters. For image-based input, computer vision detects landmarks such as wrists, elbows, and shoulders, and SMPL adjusts joint rotations until the digital body aligns with those points. For LiDAR or photogrammetry input, the entire 3D outline is used to match posture. Biological plausibility is enforced to prevent unrealistic rotations.</p><p><strong>3. Shape Adjustment<br></strong> Once the pose is matched, SMPL modifies body proportions using shape parameters (β). These parameters are statistical functions learned from thousands of body scans and represent major modes of variation such as limb length, torso depth, shoulder width, and hip size. With as few as 10 parameters, SMPL can reproduce most of the observed variation in human bodies. Real-world measurements such as waist or chest circumference can be applied directly, ensuring metric accuracy. This step links visual data from scans or photos with numerical anthropometric data.</p><p><strong>4. Error Correction<br></strong> Scans and photographs often include distortions due to clothing, reflections, shadows, or missing areas. Instead of reproducing these errors, SMPL compares the input with its statistical body model to fill gaps, smooth artifacts, and maintain realistic proportions.</p><p><strong>5. Movement Refinement<br></strong> To ensure plausible motion, SMPL applies blend shapes and skinning techniques that replicate how muscles and skin deform when joints move. This allows the digital body to bend and stretch without geometric artifacts. The model is therefore suitable for simulating how a garment behaves across a range of body motions.</p><p><strong>6. Final Output<br></strong> The final result is a 3D body model that reflects the individual’s proportions and can be posed or reset into a neutral T-pose. The neutral version is especially useful in e-commerce because it provides a consistent mannequin for garment application, sizing, and cross-brand comparison. The model at this stage is non-textured. It has accurate proportions and can be animated, but it does not include surface identifiers such as skin tone, facial features, hair, or clothing.</p><p><strong>Non-Textured Models and Privacy<br></strong> This distinction is important. Adding surface textures introduces biometric information such as facial likeness or skin color, which raises privacy and data security concerns. For applications like online shopping, non-textured models are preferred. They represent body dimensions in centimeters without disclosing identity. This enables accurate garment simulation while protecting personal data. Non-textured mannequins therefore provide a balance of precision and privacy, reducing returns and improving sizing without exposing biometric identifiers.</p><h3>How SMPL Optimizes Agentic E-Commerce</h3><p>Once the shopper’s SMPL mannequin is created, it becomes the foundation of the shopping process.</p><p><strong>Garment Files<br></strong> Clothing brands already design garments digitally using tools such as CLO3D, Marvelous Designer, or CAD software. These garment files are not just renderings; they contain pattern geometry, stitching details, and material properties such as stretch, drape, and stiffness. In an agentic e-commerce workflow, these garment files are connected directly to the shopper’s SMPL mannequin within simulation software.</p><p><strong>Virtual Draping<br></strong> When garment files are applied to the SMPL mannequin, simulation engines compute how fabrics move, fold, and stretch as the mannequin bends, walks, or raises an arm. This makes it possible to evaluate fit in terms of restriction, looseness, or sleeve and leg behavior during motion.</p><p><strong>Agent Curation and Analysis<br></strong> The agent integrates multiple sources of data to guide clothing recommendations:</p><ul><li>Fit simulations on the SMPL mannequin</li><li>Purchase and return history</li><li>Style preferences such as colors, cuts, and brands</li><li>Trend data and seasonal relevance</li></ul><p>The system filters out garments that are unlikely to fit well and highlights those that align with both physical and style requirements. It also provides explanations of fit decisions, such as whether a coat has enough space for layering or why tailored shirts are prioritized over oversized fits.</p><p><strong>Shopper Interaction<br></strong> For the shopper, the result is a digital fitting room. Instead of browsing static product pages, they view clothing directly on their mannequin. The mannequin remains neutral and non-textured to preserve privacy but is accurate enough to demonstrate fit and movement. Shoppers can rotate the model, test poses, and preview motion, reducing uncertainty about size and eliminating the need to order multiple items for comparison.</p><h3>Conclusion</h3><p>SMPL changes online shopping from estimation to precision. Instead of relying on size charts, agentic systems can test garments directly on a shopper’s mannequin, compare across brands, and recommend what will actually fit. This reduces returns, improves trust, and cuts waste at scale. For retailers, it means higher margins and fewer logistics costs. For shoppers, it means confidence in every purchase.</p><p>Several startups are already proving this transformation in action. <strong>Doji</strong> is building AI-powered avatars and virtual clothing try-ons, while <strong>Lalaland.ai</strong> generates lifelike fashion models that brands can dress with 3D garments. <strong>WearFits</strong> offers AI + AR body scanning from photos, creating precise size matches. <strong>Reactive Reality’s PICTOFiT</strong> brings realistic virtual fittings to major fashion brands through avatar simulation. <strong>GlamAI</strong> enables shoppers to upload photos and see photo-realistic try-ons, and <strong>Alter-Ego (AlterEgo.ai)</strong> develops advanced avatar infrastructure for e-commerce. These companies show that the shift toward better virtual fitting is already underway, using technology to turn smartphones into personal tailors and make online shopping more accurate, inclusive, and sustainable.</p><p>But SMPL goes far beyond retail. In healthcare, SMPL models are being used track physical change, support rehabilitation, and guide prosthetic design. In sports and fitness, they can measure progress with accuracy that current tools cannot provide. In gaming and the metaverse, they allow for consistent avatars across platforms. Finally in safety and ergonomics, they can verify fit before equipment is even produced.</p><p>The critical question is who owns this data. Body models are sensitive, and adoption will depend on control. A framework like the Self-Sovereign Anthropometric Token (SSAT) would let individuals keep their measurements in their own hands, sharing only what is required. This ensures accuracy and interoperability without compromising privacy.</p><p>The direction is clear: 3D body models will become standard across industries. SMPL provides the technical foundation, and SSAT offers the structure for safe and responsible use. Together they outline a path where digital body data becomes a trusted, everyday tool.About me</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PoxGKJNJDHV6fg_cvy7WLA.jpeg" /></figure><h3>About Me</h3><p>I’m Arav Agnihotri, a student at The Knowledge Society (TKS) exploring how AI and computer vision can be applied to anthropometrics — the science of measuring the human body. My work looks at how technologies like SMPL and markerless motion capture can be used to create accurate, privacy-preserving body models. I’m particularly focused on their potential in fashion, sports, and digital identity, where precise measurements can reduce returns, improve performance insights, and enable new forms of self-sovereign identity. I share my journey to document what I’m learning, connect with others in this space, and show how emerging tech can solve problems that affect people every day.</p><h4><strong>Lets Connect on </strong><a href="https://www.linkedin.com/authwall?trk=bf&amp;trkInfo=AQEKB_7H28ZKAQAAAZmtSewYk8mmF2clnvQXRIcbc2DH6JwD89IBWtFlZ3JfXVmGzvP9VVhOsQ4NzkXQideAcBRKRjnyOJA9Ir4EMH-PHZ-cfmcO54OfpdwUM2T4J9I-foWW9mU=&amp;original_referer=&amp;sessionRedirect=https%3A%2F%2Fwww.linkedin.com%2Fin%2Farav-agnihotri-a4910b386%3Futm_source%3Dshare%26utm_campaign%3Dshare_via%26utm_content%3Dprofile%26utm_medium%3Dios_app"><strong>LinkedIn</strong></a></h4><h4>Visit My Personal Website : <a href="https://aravs-personal-website.lovable.app/">https://aravs-personal-website.lovable.app/</a></h4><h3>Citations:</h3><p><strong>SMPL and Extensions</strong></p><ul><li>Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., &amp; Black, M. J. (2015). SMPL: A skinned multi-person linear model. <em>ACM Transactions on Graphics (TOG), 34</em>(6), 248. <a href="https://doi.org/10.1145/2816795.2818013">https://doi.org/10.1145/2816795.2818013</a></li><li>Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A. A., Tzionas, D., &amp; Black, M. J. (2019). Expressive body capture: 3D hands, face, and body from a single image. In <em>Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition</em> (pp. 10975–10985). <a href="https://doi.org/10.1109/CVPR.2019.01123">https://doi.org/10.1109/CVPR.2019.01123</a></li><li>Pons-Moll, G., Romero, J., Mahmood, N., &amp; Black, M. J. (2017). ClothCap: Seamless 4D clothing capture and retargeting. <em>ACM Transactions on Graphics (TOG), 36</em>(4), 73. <a href="https://doi.org/10.1145/3072959.3073711">https://doi.org/10.1145/3072959.3073711</a></li></ul><p><strong>Computer Vision Foundations</strong></p><ul><li>Szeliski, R. (2022). <em>Computer vision: Algorithms and applications</em> (2nd ed.). Springer. <a href="http://szeliski.org/Book/">http://szeliski.org/Book/</a></li><li>Cao, Z., Hidalgo, G., Simon, T., Wei, S. E., &amp; Sheikh, Y. (2019). OpenPose: Realtime multi-person 2D pose estimation using Part Affinity Fields. <em>IEEE Transactions on Pattern Analysis and Machine Intelligence, 43</em>(1), 172–186. <a href="https://doi.org/10.1109/TPAMI.2019.2929257">https://doi.org/10.1109/TPAMI.2019.2929257</a></li></ul><p><strong>LiDAR and Photogrammetry</strong></p><ul><li>YellowScan. (2024). LiDAR vs. photogrammetry: Key differences and applications. <em>YellowScan Knowledge Base</em>. <a href="https://www.yellowscan.com/knowledge/lidar-vs-photogrammetry-key-differences-and-applications/">https://www.yellowscan.com/knowledge/lidar-vs-photogrammetry-key-differences-and-applications/</a></li><li>Wehr, A., &amp; Lohr, U. (1999). Airborne laser scanning — An introduction and overview. <em>ISPRS Journal of Photogrammetry and Remote Sensing, 54</em>(2–3), 68–82. <a href="https://doi.org/10.1016/S0924-2716(99)00011-8">https://doi.org/10.1016/S0924-2716(99)00011-8</a></li><li>Remondino, F., &amp; El-Hakim, S. (2006). Image-based 3D modelling: A review. <em>The Photogrammetric Record, 21</em>(115), 269–291. <a href="https://doi.org/10.1111/j.1477-9730.2006.00383.x">https://doi.org/10.1111/j.1477-9730.2006.00383.x</a></li></ul><p><strong>Apple LiDAR System (Consumer Devices)</strong></p><ul><li>Apple Inc. (2020). <em>iPhone 12 Pro and iPhone 12 Pro Max: Technical specifications</em>. Apple. <a href="https://www.apple.com/iphone-12-pro/specs/">https://www.apple.com/iphone-12-pro/specs/</a></li><li>Apple Inc. (2020). <em>Use the LiDAR scanner on iPhone and iPad</em>. Apple Support. <a href="https://support.apple.com/HT211180">https://support.apple.com/HT211180</a></li></ul><p><strong>Additional Web Sources</strong></p><ul><li>Khanhha. (n.d.). <em>SMPL model introduction</em>. Retrieved October 3, 2025, from <a href="https://khanhha.github.io/posts/SMPL-model-introduction/">https://khanhha.github.io/posts/SMPL-model-introduction/</a></li><li>Max Planck Institute for Intelligent Systems. (n.d.). <em>SMPL: Skinned multi-person linear model</em>. Retrieved October 3, 2025, from <a href="https://smpl.is.tue.mpg.de/">https://smpl.is.tue.mpg.de/</a></li><li>RVS Land Surveyors. (n.d.). <em>LiDAR or photogrammetry: Which is right for your project?</em> Retrieved October 3, 2025, from <a href="https://www.rvslandsurveyors.com/lidar-or-photogrammetry">https://www.rvslandsurveyors.com/lidar-or-photogrammetry</a></li><li>Home Devices. (2023). <em>Create an Apple Home floor plan with a LiDAR sensor</em>. Retrieved October 3, 2025, from <a href="https://homedevices.app/create-an-apple-home-floor-plan-with-a-lidar-sensor/">https://homedevices.app/create-an-apple-home-floor-plan-with-a-lidar-sensor/</a></li><li>Max Planck Institute for Intelligent Systems. (n.d.). <em>SMPL code</em>. Retrieved October 3, 2025, from <a href="https://is.mpg.de/ps/code/smpl">https://is.mpg.de/ps/code/smpl</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d53d736ec470" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>