<?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 Sanveer Dhanju on Medium]]></title>
        <description><![CDATA[Stories by Sanveer Dhanju on Medium]]></description>
        <link>https://medium.com/@sanveerdhanju?source=rss-8218fc7892cb------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*Uffw9SXX1N1xJkt6.</url>
            <title>Stories by Sanveer Dhanju on Medium</title>
            <link>https://medium.com/@sanveerdhanju?source=rss-8218fc7892cb------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 15:22:50 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@sanveerdhanju/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[When One Agent Isn’t Enough: How Multi-Agent Systems Actually Work]]></title>
            <link>https://medium.com/@sanveerdhanju/when-one-agent-isnt-enough-how-multi-agent-systems-actually-work-b8190145b264?source=rss-8218fc7892cb------2</link>
            <guid isPermaLink="false">https://medium.com/p/b8190145b264</guid>
            <dc:creator><![CDATA[Sanveer Dhanju]]></dc:creator>
            <pubDate>Thu, 16 Apr 2026 02:25:33 GMT</pubDate>
            <atom:updated>2026-04-16T02:28:55.295Z</atom:updated>
            <content:encoded><![CDATA[<h4>A guide to coordinating multiple AI agents — and when you actually need to.</h4><p><a href="https://medium.com/@sanveerdhanju/how-i-think-about-langchain-a-mental-model-from-first-principles-2dc67d56c39b?postPublishedType=repub">In the last post, I introduced a simple framing: chains are closed loops and agents are open loops</a>. A closed loop means the developer defines the execution path. An open loop means the LLM defines it at runtime. That distinction organizes everything in LangChain.</p><p>But it also raises a question I didn’t address: what happens when one open loop isn’t enough?</p><p>This post is about what comes next — multi-agent systems, why they exist, when you actually need one, and the design decisions that determine whether they work or fall apart.</p><p>Here’s what this article covers:</p><ul><li>A brief timeline of how we got here</li><li>Single-agent systems vs. multi-agent systems — and how to decide</li><li>The core architecture: decompose, assign, coordinate, share state</li><li>Four coordination patterns and how to choose between them</li><li>Why managing shared context is the hardest problem</li></ul><p><strong>How we got here</strong></p><p>Every layer in the AI systems stack exists because the previous one hit a wall.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*t4T3Gsrd1Yq5sQo7ZcLr2Q.png" /></figure><p>Each layer extends the previous layer when it runs out of room.</p><p><strong>Single-agent vs. multi-agent: the decision that matters most</strong></p><p>Before touching architecture, you need to answer one question: do you actually need multiple agents?</p><p>Most of the time, the answer is no.</p><p>Simple workflows with well-crafted prompts handle 80% of real-world requirements. A single agent with the right tools is cheaper to build, easier to debug, and faster to ship. Multi-agent systems should almost never be the starting point.</p><p>Here’s the heuristic:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/950/1*dBPQGFKiumuFHlUtO__6Pw.png" /></figure><p>The key signal: if you’re giving one agent so many tools and so much context that it starts making worse decisions, you’ve hit the ceiling. That’s when you decompose.</p><p><strong>What multi-agent actually means</strong></p><p>Strip away the terminology and the core idea is straightforward:</p><p>Instead of one agent doing everything, you decompose the task across multiple specialized agents that coordinate. Each agent has a narrower scope, its own tools, and its own reasoning loop.</p><p>Through the lens of Post 1’s framing: a multi-agent system is <em>multiple open loops coordinating with each other</em>. You’re not managing one loop — you’re managing the interactions between several. Each step up adds another degree of freedom the developer has to manage.</p><p>There are four design decisions:</p><ol><li><strong>Decompose the task</strong> — what are the distinct jobs?</li><li><strong>Assign specialized agents</strong> — which agent handles which job?</li><li><strong>Choose a coordination pattern</strong> — how do agents interact?</li><li><strong>Manage shared context</strong> — what does each agent see?</li></ol><p><strong>Step 1: Decompose the task</strong></p><p>The first question is: what are the distinct jobs that need separate reasoning loops?</p><p>Take insurance claims processing after a natural disaster. A single claim requires verifying the customer’s coverage, confirming the weather event actually occurred, checking for fraud patterns, calculating the appropriate payout, and auditing the decision for compliance. Each of those is a different domain with different data sources and different reasoning requirements.</p><p>The decomposition should mirror how a team of human specialists would divide the work. If a human fraud analyst and a human coverage specialist wouldn’t be the same person, they probably shouldn’t be the same agent.</p><p><strong>Step 2: Assign specialized agents</strong></p><p>Each job becomes a specialized agent with a narrow scope: a Coverage Agent that verifies policies, a Weather Agent that confirms events, a Fraud Agent that checks for anomalies, a Payout Agent that calculates amounts, and an Audit Agent that reviews everything for compliance.</p><p>Specialization keeps each agent’s context window focused. An agent that only thinks about fraud detection will be better at it than one simultaneously trying to verify coverage, check weather, and calculate payouts. Narrower scope means fewer tools, less context, better decisions.</p><p><strong>Step 3: Choose a coordination pattern</strong></p><p>This is where it gets interesting. LangChain defines four primary coordination patterns. (Other companies like Google define architectures differently — independent, centralized, decentralized, hybrid — but the underlying tradeoffs are similar.)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mf7YCbzWAm5CAfQTOur3kQ.png" /><figcaption>3 common agent orchestration patterns</figcaption></figure><p><strong>Supervisor / Orchestrator</strong> — one central agent delegates tasks to specialized workers and synthesizes their outputs.</p><p><strong>Handoffs / Swarm</strong> — agents pass control directly to each other with no central coordinator, each deciding the next step.</p><p><strong>Skills</strong> — an orchestrator invokes specialized capabilities that execute and return, without making independent decisions.</p><p><strong>Router</strong> — a classifier routes the entire request to a single specialist agent, with no coordination between agents.</p><p>Not every pattern works for every requirement. Here’s a scoring framework for evaluating them across four criteria:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*e4VN7nJoMH2qrJtpj4FOrA.png" /><figcaption>Criteria and scores provided by LangChain Youtube video</figcaption></figure><p><strong><em>Distributed development:</em></strong><em> can different teams maintain components independently? <br></em><strong><em>Parallelization:</em></strong><em> can multiple agents execute concurrently? <br></em><strong><em>Multihop:</em></strong><em> does the pattern support calling multiple sub-agents in series? <br></em><strong><em>Direct user interaction:</em></strong><em> can sub-agents converse directly with the user?</em></p><p>Choosing a pattern isn’t about finding the best one — it’s about understanding which tradeoffs your product can absorb.</p><p><strong>Step 4: Managing shared context — the hardest problem</strong></p><p>This is where most multi-agent systems break. Each agent runs its own loop and builds its own understanding of the task. The question becomes: what does each agent need to see?</p><p>There are three approaches:</p><p><strong>Shared state</strong> — agents read from and write to a common state object. Every agent can see what other agents have produced. Most flexible, but expensive — multi-agent systems can cost five to ten times as much as single agents because every agent processes the full context. It also adds noise: the fraud agent doesn’t need weather verification details.</p><p><strong>Tool access</strong> — agents can’t see each other’s state, but they can call each other as tools. The Coverage Agent becomes a callable function that returns a structured verdict: covered or not covered, with key conditions. Downstream agents get the decision without the reasoning trace. Cleaner boundaries, but less flexible.</p><p><strong>No shared context</strong> — each agent operates in complete isolation, seeing only its own inputs and outputs. Cheapest, but brittle. If the Coverage Agent discovers the policy was cancelled, the Payout Agent needs to know before calculating a number. Without shared context, agents make decisions on incomplete information.</p><p>The right answer depends on your product. But in every case, deciding what flows between agents is a product design question, not an engineering one. The PM who understands multi-agent architecture is the PM who can define these context boundaries.</p><p>But knowing the architecture doesn’t mean it works in production. Agents fail in specific, predictable ways — and multi-agent systems introduce entirely new failure modes. A single loop can wander or hallucinate. Multiple coordinated loops can do all of that, plus conflict with each other, duplicate work, or cascade failures across loop boundaries.</p><p>Next piece is where architecture meets reality: where these systems break, and the production patterns that fix them.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b8190145b264" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How I Think About LangChain: A Mental Model From First Principles]]></title>
            <link>https://medium.com/@sanveerdhanju/how-i-think-about-langchain-a-mental-model-from-first-principles-2dc67d56c39b?source=rss-8218fc7892cb------2</link>
            <guid isPermaLink="false">https://medium.com/p/2dc67d56c39b</guid>
            <dc:creator><![CDATA[Sanveer Dhanju]]></dc:creator>
            <pubDate>Wed, 15 Apr 2026 02:08:58 GMT</pubDate>
            <atom:updated>2026-04-16T02:27:46.932Z</atom:updated>
            <content:encoded><![CDATA[<h4>A conceptual guide to chains, agents, and LangGraph — for PMs, founders, and anyone who doesn’t want to start with the docs.</h4><p>Here’s what I found: most tutorials jump straight into classes and API calls. Very few explain the <em>conceptual architecture</em> — the mental model that makes everything else click. So I’m going to share what I landed on, written for PMs, founders, and anyone who needs to reason about these systems without living in the codebase.</p><p>Because once you have the right mental model, LangChain is surprisingly simple.</p><p>Here’s what this article covers:</p><ul><li>Why LangChain exists — a brief history</li><li>The atomic building block: the LLM Chain</li><li>Six chain patterns that became design foundations</li><li>LCEL: the modern syntax that replaced them</li><li>The chain vs. agent distinction — the most important concept in the framework</li><li>Three agent types and who controls the loop</li><li>LangGraph: the orchestration layer that ties it all together</li></ul><p><strong>A brief history, because it matters</strong></p><p>LLMs arrived and gave us something remarkable — incredibly smart question-and-answer machines. For the first time, you could have a natural conversation with software and get genuinely useful responses. But they were just that: conversational. They could <em>talk</em>, but they couldn’t <em>do</em> anything. They couldn’t search the web, query a database, call an API, or take an action in the real world.</p><p>LangChain emerged to solve that problem. It’s a framework that sits on top of LLMs and gives them the ability to do things — to connect with tools, data sources, and multi-step reasoning workflows. As developers started building with it, common patterns kept showing up. Certain ways of wiring components together proved useful again and again, and those patterns got named and formalized.</p><p>The framework has evolved since then. The specific chain types that were formalized early on have been replaced by more flexible syntax. But the patterns themselves are still the right way to understand what this framework is capable of and how to think about designing with it.</p><p><strong>Two layers of capability</strong></p><p>For the purpose of building a mental model, there are two layers worth understanding first:</p><p><strong>LangChain</strong> gives you the building blocks — chains, tools, prompts, parsers. Think of these as the Lego bricks.</p><p><strong>LangGraph</strong> sits on top and gives you the orchestration layer — the ability to wire those blocks into stateful, dynamic systems. Think of this as the blueprint.</p><p>Let me walk through each.</p><p><strong>LangChain: building with Lego blocks</strong></p><p>The atomic unit is something everyone already knows — an LLM. You give it input, it gives you output.</p><p>Now bundle that LLM with a prompt template that standardizes the input and an output parser that standardizes the output. What you get is called an <strong>LLM Chain</strong>. This is the primitive building block — the single Lego brick that everything else is built from.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AfaOaFC2RH90_229PEByvw.png" /><figcaption>LLM Chain is the Lego Brick</figcaption></figure><blockquote>Here’s a simple example: you give the system a news article about Tesla and ask “what’s the sentiment of this article?”</blockquote><ol><li>The <strong>prompt template</strong> shapes the question,</li><li>The <strong>LLM</strong> analyzes it</li><li>The<strong> output parser</strong> returns a structured result — positive, negative, or neutral.</li></ol><p>The interesting part is what happens when you start linking these bricks together. Different ways of connecting LLM Chains give rise to different patterns, and six of them showed up often enough to get named:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/886/1*xelhLUAR07pZrMudWxbrbQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/970/1*J7I_WQKz6iana7wzlo1Afg.png" /></figure><p>Take our Tesla example. A <strong>Sequential Chain</strong> would:</p><ol><li>Search for recent Tesla news</li><li>Then summarize what it finds,</li><li>Then analyze the sentiment</li></ol><p>A<strong> Router Chain </strong>would:</p><ol><li>Look at the sentiment result</li><li>Then route to different report formats depending on whether it’s positive, negative, or mixed.</li></ol><p>A <strong>Map-Reduce Chain</strong> would take ten articles, send each to its own summarization chain in parallel, then combine all ten summaries into one overall picture. Each pattern is a different way of wiring the same Lego bricks.</p><p>A <strong>Refine Chain</strong> would take those same ten articles but process them sequentially — drafting a sentiment summary from the first article, then refining it with each subsequent article, one at a time, until the final summary reflects all ten.</p><p>You don’t need to memorize these. But you do need to understand them — because they represent the fundamental design patterns for how LLM-powered pipelines can be structured. Routing, parallelization, iterative refinement, sequential handoff. These ideas don’t go out of date.</p><p>Today, modern LangChain replaces all six with a more flexible syntax called <strong>LCEL</strong> (LangChain Expression Language), which lets you compose the same patterns using just four primitives:</p><p>Same patterns, cleaner syntax.</p><p><strong>Chains vs. Agents: the distinction that organizes everything</strong></p><p>Everything we’ve covered so far — the chain types, LCEL — is <strong>deterministi</strong>c <strong>closed loop </strong>systems: the developer defines the execution path before any input arrives, and the LLM fills in content but never decides structure.</p><p>An <strong>agent</strong> is what happens when you have an <strong>open loop</strong> system.</p><p>Instead of a fixed path,</p><blockquote>an agent has a <strong>core loop</strong>: think, act, observe, repeat.</blockquote><p>Back to our example.</p><p>Now give an agent a harder version of the same task: “Research Tesla’s current market position, find any recent controversies, and give me a sentiment-weighted summary I can send to my team.” The agent might search Google first, decide the results are too old, try a financial news API instead, realize it needs earnings data for context, pull that, discover a recent lawsuit it hadn’t anticipated, research that separately, then synthesize everything into a single report. It decided the path as it went. No two runs would look the same.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kvCCEHAdEIZ90ZMBojn5-w.png" /><figcaption>Agents have access to modules and decide when to use each</figcaption></figure><p>That’s the single most important distinction in this entire framework. <strong>Closed loop</strong> vs.<strong> open loop.</strong> Chain vs. agent. Once you see it, every other concept in LangChain is just a variation on one side or the other.</p><p><strong>Three agent types worth understanding</strong></p><p>Just like the chain side, the agent side has named patterns that became foundational.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5y6bO6Hipc1Ckcgqxe_Qew.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GTzlNwU6hYWtdDey-5XCtA.png" /><figcaption>A Functions Agent is similar to a ReAct agent, it but does not think aloud.</figcaption></figure><p><strong>ReAct Agent</strong> — an agent core loop + printed reasoning steps + tools</p><p><strong>Functions Agent</strong> — an agent core loop + tools + structured output</p><p><strong>Plan-and-Execute Agent</strong> — an agent core loop split in into two phases: a planner LLM creates a step-by-step plan, then an executor LLM works through it sequentially</p><p>All three of these agent patterns share a limitation — the developer has limited control over the exact shape of the loop. You can choose the pattern, but you can’t fine-tune the topology. <strong>That’s the gap LangGraph was built to fill.</strong></p><p><strong>LangGraph: the orchestration layer</strong></p><p>LangGraph sits on the agent side of the line, but it gives you back developer control over the shape of the system. It’s the answer to a specific problem: how do you get the flexibility of an open loop without losing the predictability of a closed one?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wPfpG5FNg4rXHZTXcx33bA.png" /></figure><p>Conceptually, LangGraph takes the same Lego blocks — your LLM Chains, your tools, your logic — and lets you wire them into an explicit graph. Instead of a linear pipeline, you define <strong>nodes</strong> (where work happens) and <strong>edges</strong> (how work flows between them, including conditional routing based on what the system has learned so far). You decide which nodes exist, how they connect, and under what conditions the system moves between them. The LLM drives reasoning within each node, but the structure is yours.</p><p>For our Tesla example, you’d define the graph explicitly: a news search node, an earnings node, a legal research node, a summarization node, and a sentiment analysis node. You’d define the edges between them — including conditional edges like “if the news results mention a lawsuit, route to the legal research node before summarizing” and “if the search results are older than 48 hours, go to the earnings node for more current data.” The LLM does the thinking within each step. You control the map.</p><p>The key new concept LangGraph introduces is <strong>state</strong> — a shared object that every node in the graph can read from and write to as execution progresses. State tracks everything: which sources have been searched, what articles have been found, what the current sentiment score looks like, whether a controversy was detected, whether the system has enough information to generate a final report or needs to keep digging. It’s what gives the graph a sense of “where am I and what do I know so far?”</p><p>This is different from <strong>memory</strong>, which is just conversation history — it answers “what did we say before?” State is the superset. Memory is one field inside it. That distinction matters more than most tutorials let on.</p><p><strong>The complete mental model</strong></p><p>Here’s the full arc:</p><p>LLMs gave us smart conversational machines. LangChain gave those machines the ability to act. Developers started wiring LLM Chains together, and common patterns emerged — six chain types that became design foundations. LCEL replaced the syntax but preserved the patterns. All of that lives on the deterministic side: closed-loop systems where the developer draws the flowchart.</p><p>Agents opened the loop. The LLM started drawing its own flowchart at runtime. Three agent patterns emerged — ReAct, OpenAI Functions, Plan-and-Execute — each representing a different answer to the question of who controls the loop.</p><p>LangGraph brought it all together: an orchestration layer where the developer defines the graph topology explicitly, the LLM reasons within each node, and a shared state object tracks everything the system knows as it executes.</p><p>That’s the mental model. Everything in LangChain sits somewhere in that stack.</p><p>I’m sharing this series because I think the AI systems layer is something technical PMs and founders need to genuinely understand right now — not just reference in a strategy deck. This is how I’d explain it to my team.</p><p>Next post: <a href="https://medium.com/@sanveerdhanju/when-one-agent-isnt-enough-how-multi-agent-systems-actually-work-b8190145b264">what happens when a single agent isn’t enough — a deep dive into multi-agent architectures.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2dc67d56c39b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Nutrition Economics at McDonald’s]]></title>
            <link>https://medium.com/@sanveerdhanju/nutrition-economics-at-mcdonalds-763ed58f356f?source=rss-8218fc7892cb------2</link>
            <guid isPermaLink="false">https://medium.com/p/763ed58f356f</guid>
            <category><![CDATA[health]]></category>
            <category><![CDATA[mcdonalds]]></category>
            <category><![CDATA[nutrition]]></category>
            <category><![CDATA[freakonomics]]></category>
            <dc:creator><![CDATA[Sanveer Dhanju]]></dc:creator>
            <pubDate>Sun, 29 Oct 2017 09:10:41 GMT</pubDate>
            <atom:updated>2017-11-01T14:13:19.067Z</atom:updated>
            <content:encoded><![CDATA[<p>McDonald’s has long been heralded as inexpensive food that is priced for value. With the popularization of the ‘dollar menu’, <strong>I was curious to see what McDonald’s menu item provides the most nutritional value for my money. In pursuit of this, I took a look at nutritional variance globally </strong>to determine if burgers offered the same value around the world. Next, focusing in on Canada’s menu, I looked at the <strong>proportion of energy derivatives </strong>to assess the quality of calories. Third, I looked at the same menu items with a <strong>ratio of cost to nutrition</strong> to see how the cost of calories compared across sandwiches. Combining these three analysis allowed me to determine what item is the best bang for your buck at McDonald’s.</p><p><strong><em>Are all burgers created equal?</em></strong></p><p>I took a look at the calories in a Big Mac in various countries around the world to see if a dollar could buy you the same amount of energy around the world.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2xUiJ9lV8T7J2O5Ee5EorA.png" /></figure><p>Assuming the McDonald’s methodology for measuring nutritional information is standardized around the world, there are quite a few differences in Big Macs globally. It is common practice for restaurant chains to alter their menu offerings to adapt to local tastes and cuisine. As a result, <strong>the exact composition of sandwiches are not comparable around the world.</strong></p><p>Due to the limited availability of information on international nutrition facts, only a sample set of countries could be used in this analysis. Across the sampled countries, the calories in a Big Mac varied by as much as up to 15%. The Big Mac seems to be more calorie dense in North America then other regions in the world. Interestingly, Saudi Arabia had the highest calories in a Big Mac at 570 calories. This is much higher than the most Asian regions with lighter calorie sandwiches.</p><p><strong>Nutritional information from the Canadian menu was sourced for the following analysis.</strong> Canada lies in the middle of the spectrum of the sampled countries and is a good representation of McDonald’s operating in a competitive market.</p><p><strong><em>Are all calories created equal?</em></strong></p><p>The<strong> three main sources of calories in a diet are proteins, carbohydrates, and fats</strong>.<strong> </strong>It is important to balance these sources of energy, which also provide insight into the quality of ingredients in a meal. Generally, there are 4 calories / g protein, 4 calories / g of carbohydrates, and 9 calories / g of fat. I took a look at the distribution of calories across these three sources for several popular menu items to understand the type of energy I was buying. Menu items are compared within categories of beef, chicken, and salad. However, it is important to note that although the calorie count was considered, there are important differences in the types of proteins, carbs, and fats that were not taken into account for simplicity.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hlxdrN0i_DWMFdVEEMJgZw.png" /></figure><p>The source of energy for menu items varies across the board, while roughly 20% of calories are from proteins, 50% are form fats, and 30% are from carbohydrates. Most of the energy from popular items like chicken nuggets, and the McChicken come from fats. On the other hand, cheaper items such as the McTaster, cheeseburger, and McDouble have high amounts of energy derived from carbohydrates. Likely due to their smaller size, the breading of these items account for more of the total mass and caloric source.</p><p>The items with the highest proportions of protein are the salads and beef items, while the chicken items have the lowest proportion of proteins. Interestingly, the grilled chicken salad had twice the proportion of protein than the grilled chicken. The batter and frying of the chicken significantly increase the calorie count from fats.</p><p>Among the sandwiches, the Angus burger, McDouble, and Double Big Mac stand out in their protein composition. However, the McDouble has much less calories derived from fats, and more from carbohydrates.<strong> In terms of calorie derivatives, the McDouble stands out for it’s composition of high protein and low fat</strong>.</p><p><strong><em>Do all calories cost the same?</em></strong></p><p>To consider the true value of these items, I calculated and compared the cost per gram of protein and cost per calorie of these items.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*W-omOkqiZ-N1mNByzFksMA.png" /></figure><p>The first note of interest is that beef items are much cheaper per calorie and gram of protein consumed. This makes sense as the cost of procuring chicken ingredients is higher than beef, which is reflected in the value of these items. By purchasing more beef items, you will be getting more protein and calories for every dollar spent.</p><p>Secondly, the ‘value’ items, such as the Junior Chicken, McDouble, and Cheeseburger, offer the most monetary value in terms of protein and energy. These items are at the low end of the spectrum and provide cheap fuel. It seems that the sandwiches with larger serving sizes do not give you more value for your money. These larger menu items have a premium, especially among the larger chicken items.</p><p>Third, the salads seem to be much more cost effective on a nutrition basis than sandwiches with chicken. Although the price of salads are higher per calorie, they sell at a similar protein cost point as beef.</p><blockquote><strong>So what’s the verdict?</strong></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/448/1*fZ7qqRfKnxL9zTYAA5JSlA.png" /></figure><p>The McDouble is a clear winner for true value at McDonald’s. The composition of calories in the McDouble rival the crispy chicken salad, at the lowest cost point per calorie across the board. So, the next time you are feeling hungry and have a light wallet, consider a McDouble for your next fast food run.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=763ed58f356f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Rise of 3D Printed Drugs]]></title>
            <link>https://medium.com/@sanveerdhanju/rise-of-3d-printed-drugs-8fc5c21c36ae?source=rss-8218fc7892cb------2</link>
            <guid isPermaLink="false">https://medium.com/p/8fc5c21c36ae</guid>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[pharmaceutical]]></category>
            <category><![CDATA[3d-printing]]></category>
            <category><![CDATA[strategy]]></category>
            <dc:creator><![CDATA[Sanveer Dhanju]]></dc:creator>
            <pubDate>Sun, 20 Aug 2017 22:29:54 GMT</pubDate>
            <atom:updated>2017-10-29T09:30:43.913Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aCInKiMs1CJBCg15JfE4Rg.png" /></figure><p><strong>Rise of 3D Printed Drugs</strong></p><p>3D Printing (3DP) has permeated many fields of medicine, pushing new innovations and possibilities. Recently developments in 3DP technology have enabled its implementation in printing pharmaceuticals. Startups are emerging with promising advancements in personalized medicine and distribution of medicine.</p><p><em>Aprecia Pharmaceuticals</em></p><p>Based out of Pennsylvania, Aprecia Pharmaceuticals is a commercial stage startup in specialty pharmaceuticals. They develop and manufacture products using their proprietary ZipDose technology. In 2015, Aprecia Pharmaceuticals became the first company to have a 3DP prescription pill approved by the FDA. Spritam is a generic form of levetiracetam, a treatment for epilepsy, that is 3DP with ZipDose. ZipDose uses 3DP to bind together several thin layers of powdered drug that is interspread with liquid droplets which make the pill effortless to swallow for those with difficulty swallowing.</p><p>Aprecia’s strategy is to repackage highly prescribed, highly dosed drugs and make its delivery more effective and convenient. With a total of 4 drugs in their pipeline, Aprecia is expanding its portfolio of drugs with large market potential. Aprecia is currently positioned as a potential major threat to generic drug manufacturers.</p><p><em>Teva Pharmaceuticals</em></p><p>Teva Pharmaceuticals is a multinational pharmaceutical based in Petah Tikva, Israel. With specialties in developing, manufacturing, and marketing generic drugs, Teva is currently the largest generic drug manufacturer in the world. It has over 70 manufacturing sites in Israel, North America, Europe, and South America, and it’s 2015 gross revenues totalled $19.5 billion. Teva’s growth strategy is focussed on building a stronger brand by including more high impact products in their portfolio. This focus in the generics market is illustrated by their 2015 acquisition of Actavis Generics, the second largest pharmaceutical division in generic drugs. However, simply acquiring more generics will not be enough to sustain its growth in the coming age of 3DP drugs.</p><p><strong>The 3DP High</strong></p><p>3DP technology holds significant implications in the pharmaceutical industry. The most substantial advantages include:</p><blockquote>1. Drug personalization</blockquote><blockquote>2. Distribution</blockquote><blockquote>3. Cost</blockquote><p>Printing pills is a much more flexible process in which prescriptions can be manipulated to tailor ingredients and form to specific patients. Traditionally, generic products are undifferentiated and emerge as soon as branded drugs patent expire. This results in many generic pharmaceuticals producing the same product, where the lowest cost product would emerge on top. However, with 3DP advancements generic products can now be differentiated through innovations in their production and mechanism. Furthermore, these innovations can be patented to secure a competitive advantage in the generic drug industry.</p><p>Distribution of prescription drugs involves a complex supply chain with many intermediates. 3DP allows for the possibility of a decentralized supply network where becomes localized rather than centralized. It would change how drugs are manufactured and delivered affecting the speed, risk, and quality of the supply chain.</p><p>With advancements in 3D printing technology, materials and equipment become cheaper. This enables lower cost drugs with higher margins for pharmacy’s.</p><p><strong>Challenge for Teva</strong></p><p>The biggest challenge for Teva Pharmaceuticals with the advent of this technology is twofold:</p><blockquote>1) Can’t compete with generic drug innovation</blockquote><blockquote>2) Losing control over the supply chain</blockquote><p><em>Innovations</em></p><p>Startups like Aprecia are vetting popular generic drugs from the libraries of large generic producers like Teva. Teva’s market share in these products will become threatened as 3DP producers have more value add in the product, and could possibly offer it at a lower cost. How generic producers compete will no longer be based on only the lowest cost producer, but also on who holds patents on design innovations. However, the pharmaceutical industry is very sensitive as it is directly tied to the health of people. Hence, 3DP innovations in drugs are not yet legitimized as a safe alternative.</p><p>3DP democratizes the production of drugs, allowing for pharmacies and individuals to create their own drugs. Traditionally manufacturers pass products on to distributors, which include importers and wholesalers. The drugs are then distributed to dispensaries and retailers, and finally to the patient.</p><p>As startups and pharmacies gain the convenience and flexibility to create their own generic drugs, the relevance of traditional generic drug manufacturers like Teva is threatened. Taking a close look at the supply chain reveals stress points that could facilitate its disruption.</p><p><em>The Supply Chain</em></p><p>The supply chain is complex and requires substantial coordination. Responsibility for inventory management and demand planning falls on the dispensaries, which can get complicated. These factors create friction in the supply chain and creates an incentive for pharmacies to adopt a more convenient solution. 3DP offers the convenience of on demand supply, product innovation, and control into the hands of players downstream the supply chain. Thus, the supply chain is prime for disruption in the near future. However, the current barrier to this fate is the difficulty for regulators to establish 3DP guidelines.</p><p>Teva’s top five generic drugs makeup $2 billion of its revenues, where 50% of these revenues are from the United States. Even a modest loss in US market share of these few products would be a substantial loss for Teva. Losing 5% of this market translates to $50 million in lost revenue. With the emergence of more 3DP startups targeting its broad portfolio of generics, the impact could be much greater.</p><p><strong>Strategy</strong></p><p>To stay relevant, Teva pharmaceuticals should develop a network of decentralized 3DP facilities, and work with regulators to develop 3DP guidelines for pharmaceuticals.</p><p><em>New Tech</em></p><p>Teva has expressed interest in 3D printing, but has yet to aggressively pursue it. It has a history of expanding its generic portfolio by acquiring products from other businesses. Teva pharmaceutical should now acquire Aprecia Pharmaceuticals to acquire their patents on 3DP drug and ZipDose technology. These two capabilities will allow Teva to bring many of it’s other generic products to market quickly. This would secure an innovation that differentiates Teva from its competition that cannot be copied. Starting by digitizing the drugs that would most benefit form ZipDose, Teva can increase its market share in those markets, and use the funds to invest into R&amp;D for new 3DP innovations.</p><p><em>New Model</em></p><p>Teva has a massive supply network connected to many pharmacies around the world. This scale can be leveraged to implement 3DP into its own supply chain. The largest point of friction they face within their supply chain is between the distributor and dispensary. By building a network of 3DP facilities across the states, Teva can bring in a localized production model. Facilities would house 3D printing machines and take orders from local pharmacies with detailed specifications. Completed drugs could then be transported locally to dispensaries quicker and cheaper than traditional methods. This would retain intellectual property in the hands of Teva while servicing the pharmaceutical community. If the 3DP technology and chemical blueprints are secured, the issue of piracy by third parties should be contained.</p><p>A large barrier to pharmacies printing their own drugs instead is regulation. Although a 3DP drug has been approved by the FDA, regulating the 3DP process of drugs is uncharted territory. Regulators need to ensure that printed drugs are made consistent, tamper proof, and error free. Regulating printing for individual pharmacies is difficult feat and infeasible as every pharmacy would have to be individually checked. It would be much easier for regulators to approve a standardized approach through 3DP facilities. Teva should collaborate with the FDA to develop regulations for 3DP facilities. Regulating this will help the public accept the safety of 3DP drugs and legitimize it as a suitable innovation.</p><p><strong>Impact</strong></p><p><em>Drug innovation</em></p><p>Approximately 200 000 people die each year from adverse drug reactions (ADR). 80% of these ADRs are dose dependent. This implies that better dose control and personalization can reduce the prevalence of ADRs. 3DP allows for more flexibility in tailoring drug ingredients and delivery mechanism. About 40% of adult Americans have difficulty swallowing a pill. Innovations like the ZipDose technology have huge potential to improve the quality of life for many people.</p><p><em>Reach</em></p><p>With a network of 3DP facilities, Teva’s reach would be expanded. Currently a large issue facing developing nations in access to drugs is distribution and price. This strategy would solve the distribution issue as the supply chain would no longer be dependent on international distributers, but only on local distribution.</p><p><em>Profitability</em></p><p>Teva is currently the market leader in generic drugs, with 18% of the $80 billion market in the US. 3DP can allow Teva to capture more of the US market share through product innovations, and can reduce distribution costs that increase margins. Distributor margins can range from 10–30% and make up a significant portion of cost. Furthermore, a professor at the University of Glasgow is developing a 3D printer for drugs for as low as $1200. This illustrates how low the upfront capital investment could be. This solution would increase the margin for Teva while improving patient outcomes.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8fc5c21c36ae" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>