<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Puran Water&apos;s Blog</title>
    <description>AI-native process engineering for industrial wastewater and waste-to-value projects. Combining traditional engineering principles with cutting-edge technology solutions.</description>
    <link>https://puranwater.com</link>
    <atom:link rel="self" type="application/rss+xml" href="https://puranwater.com/rss.xml" />
    <language>en-us</language>
    <lastBuildDate>Wed, 15 Jul 2026 23:30:42 GMT</lastBuildDate>
    <item>
      <title>The AI Paradox in EPC - Why Agentic Workflows Will Create More Engineering Jobs, Not Fewer</title>
      <link>https://puranwater.com/ai-paradox-epc-agentic-workflows-engineering-talent</link>
      <guid isPermaLink="true">https://puranwater.com/ai-paradox-epc-agentic-workflows-engineering-talent</guid>
      <description>How AI-driven project design and execution will unlock massive pent-up demand for capital projects, creating a Jevons paradox where engineering talent becomes more valuable than ever</description>
      <pubDate>Thu, 12 Jun 2025 00:00:00 GMT</pubDate>
      <content:encoded><![CDATA[<p>One of the ongoing debates I have been following is the debate on the impact AI will have on the white-collar job market. Several prominent voices within the leading labs are warning of severe and rapid dislocation, among them <a href="https://www.axios.com/2025/05/28/ai-jobs-white-collar-unemployment-anthropic">Dario Amodei</a> (CEO of Anthropic). As someone who has tried to use AI in my own workflow, it seems clear to me that even in its current form, the models we have access to are a significant productivity multiplier when used well. But that does not mean the demand for human labor (i.e. my time) has decreased. In fact, it has been the opposite: the list of things I can do with AI that were inaccessible before, when I did not have the equivalent of a software engineer, business consultant, and technical consultant at my disposal, has grown so long that I find myself (paradoxically) with <strong>more</strong> work to do. The nature of the work is different, but the same dynamic could play out across the job market — skills will be displaced, but human labor will not. There will still be demand for people who can orchestrate AI agents toward a vision, use judgment to course correct, and sell what is produced to others.</p>
<h2>Jevons Paradox</h2>
<p>Jevons Paradox <a href="https://en.wikipedia.org/wiki/Jevons_paradox">&quot;occurs when technological advancements make a resource more efficient to use… however, as the cost of using the resource drops, if the price is highly elastic, this results in overall demand increasing, causing total resource consumption to rise.&quot;</a> </p>
<p>I believe the case for this phenomenon occurring in the EPC business as agentic engineering workflows diffuse through the industry is compelling.
<a href="https://towardsdatascience.com/the-iron-law-of-megaprojects-18b886590f0b/">Oxford professor Bent Flyvbjerg&#39;s extensive research</a> documents what he calls the &quot;Iron Law of Megaprojects&quot;: over budget, over time, under benefits, over and over again. The statistics are sobering:</p>
<ul>
<li><strong>Public Infrastructure</strong>: Average cost overruns of 20-80%, with some projects exceeding 400%</li>
<li><strong>Private Sector</strong>: <a href="https://www.offshore-energy.biz/ey-oil-gas-megaprojects-exceed-budgets-cost-overruns-at-500b/">Oil &amp; gas megaprojects alone have accumulated $500 billion in cost overruns</a></li>
<li><strong>Schedule Performance</strong>: 70% of projects experience significant delays</li>
<li><strong>Benefit Realization</strong>: Actual returns typically achieve only 60-70% of projected values</li>
</ul>
<p>This chronic underperformance has created a vicious cycle. As investors and governments factor in likely overruns, they:</p>
<ul>
<li>Add 30-50% contingency buffers to budgets</li>
<li>Raise hurdle rates for project approval</li>
<li>Shelve marginal but socially valuable projects</li>
<li>Slow infrastructure modernization globally</li>
</ul>
<p>The result is massive accumulation of unmet demand for capital projects worldwide.
Using <a href="https://chatgpt.com/share/684c4707-6ec8-8007-83df-9b85be018e82">ChatGPT&#39;s Deep Research</a>, I attempted to quantify the value of this &quot;pent-up demand&quot; and estimate how increased efficiency in delivering projects (and increased reliability of cost and schedule estimates) could &quot;pull forward&quot; this demand and increase the labor requirement despite a lower labor intensity per-project.</p>
<h3>Infrastructure Gap</h3>
<ul>
<li><strong>Current shortfall</strong>: <a href="https://outlook.gihub.org/">$15-20 trillion through 2040</a> (business-as-usual scenario)</li>
<li><strong>True need</strong>: Likely far higher to bring global infrastructure to developed economy median standards</li>
<li><strong>Annual investment required</strong>: Would need to double from $3.3 to $4+ trillion to close both the current shortfall and meet the demand for modern infrastructure</li>
</ul>
<p><strong>[Infrastructure Gap Timeline]</strong></p>
<p>The global infrastructure investment shortfall through 2040 — roughly $15–20 trillion in baseline-case underbuild.</p>
<h3>Private Sector Capital Gap</h3>
<ul>
<li><strong>Industrial capacity shortfall</strong>: The private capital stock needed to meet global consumption at developed market levels likely represents multiples of the existing global stock</li>
<li><strong>Sectoral variations</strong>: Elastic sectors (electronics, pharma) vs inelastic sectors (utilities, basic materials)</li>
</ul>
<h3>The Dual Impact of AI on Project Economics</h3>
<p><strong>1. Labor Efficiency (The Obvious Impact)</strong></p>
<ul>
<li>Engineering and project management hours reduced by ~25%</li>
<li>But these services represent only 15-25% of Total Installed Cost (TIC)</li>
<li>Net impact on project cost: ~5% reduction</li>
</ul>
<p><strong>2. Decision Intelligence - AI enables fundamentally better decisions across the project lifecycle:</strong></p>
<ul>
<li>Design Optimization: AI can explore dozens of design alternatives vs. just a few manually, optimizing equipment sizing and configuration (assume 3-7% TIC reduction)</li>
<li>Procurement Intelligence: Global supplier matching, price forecasting, and logistics optimization affecting the 40-60% equipment/materials portion of TIC (assume 5-10% TIC reduction)</li>
<li>Schedule Optimization: Monte Carlo simulations and weather-adjusted scheduling reduce schedule slippage and assosiated costs (interest during construciton, contractor mobilization expenses) (assume 2-5% TIC reduction)</li>
<li>Risk Mitigation: Better decision making reduces contingency needs (assume 2-5% TIC reduction)</li>
</ul>
<p>Combined, AI-driven decision intelligence could reduce TIC by 12-27% - far exceeding the labor efficiency gains.</p>
<h2>The Real Mechanism: It&#39;s Not About Demand Elasticity</h2>
<p>The traditional Jevons Paradox assumes high demand elasticity. But infrastructure demand is actually <strong>inelastic</strong> (0.06-0.52) - people need water, power, and roads regardless of price. So how does the paradox still apply?</p>
<h3>Understanding the Dual Mechanism</h3>
<p><strong>For Infrastructure Projects:</strong></p>
<ul>
<li>The bottleneck isn&#39;t end-user demand but <strong>project bankability</strong></li>
<li>Currently 30-40% of needed projects fail investment thresholds due to execution risk</li>
<li>AI doesn&#39;t make demand elastic - it makes more projects investable</li>
</ul>
<p><strong>For Private Industrial CAPEX:</strong></p>
<ul>
<li>Consumer goods often have elastic demand (&gt;1.0)</li>
<li>AI creates a double effect: production efficiency AND investment de-risking</li>
<li>Competitive dynamics force investment to capture market share</li>
</ul>
<h3>How AI Expands the Project Universe</h3>
<p><strong>1. Risk Reduction (Primary for Infrastructure)</strong></p>
<ul>
<li>Reduces cost overruns: 20-45% → 5-10%</li>
<li>Improves revenue forecasting: ±20% → ±5% error</li>
<li>Lowers WACC by ~100bps through reduced uncertainty</li>
<li><strong>Result</strong>: 15-20% of &quot;unbankable&quot; projects become viable</li>
</ul>
<p><strong>2. Volume Expansion (Primary for Private CAPEX)</strong></p>
<ul>
<li>10% manufacturing cost reduction → 8% price drop</li>
<li>With elasticity &gt;1: volume increases 10-15%</li>
<li>Higher throughput justifies new capacity</li>
<li><strong>Result</strong>: 15-30% more projects in elastic sectors</li>
</ul>
<p><strong>3. Decision Intelligence (Both Sectors)</strong></p>
<ul>
<li><strong>Design Optimization</strong>: AI explores dozens more alternatives (3-7% cost savings)</li>
<li><strong>Procurement Intelligence</strong>: Global supplier matching, price forecasting (5-10% savings)</li>
<li><strong>Schedule Optimization</strong>: Weather-adjusted, risk-based scheduling, lowers delay related costs (interest during construction, contractor mobilization) (2-5% savings)</li>
<li><strong>Risk Mitigation</strong>: Better decision making and lower variance from proforma to actual costs reduces need for contingency (2-5% savings)</li>
</ul>
<h3>The Sectoral Jevons Paradox</h3>
<p>The paradox strength varies dramatically by sector:</p>
<table>
<thead>
<tr>
<th>Sector Type</th>
<th>Project Volume Increase</th>
<th>Engineering Hours/Project</th>
<th>Net Engineering Demand</th>
</tr>
</thead>
<tbody><tr>
<td><strong>Infrastructure</strong></td>
<td>+15-20% (bankability)</td>
<td>-25%</td>
<td><strong>+5-10%</strong></td>
</tr>
<tr>
<td><strong>Private CAPEX (Elastic)</strong></td>
<td>+25-35% (volume + bankability)</td>
<td>-25%</td>
<td><strong>+15-30%</strong></td>
</tr>
<tr>
<td><strong>Private CAPEX (Inelastic)</strong></td>
<td>+10-15% (bankability only)</td>
<td>-25%</td>
<td><strong>0-10%</strong></td>
</tr>
</tbody></table>
<p><strong>[Jevons Paradox Calculator]</strong></p>
<p>Models how AI productivity gains in engineering can expand the project pipeline faster than they shrink per-project hours, so demand for engineers grows rather than shrinks.</p>
<p>While demand for labor may increase, the transition won&#39;t be smooth. Traditional engineering skills that can be automated (routine calculations, standard designs, basic documentation) will give way to roles requiring AI orchestration, judgment, and complex decision-making.</p>
<p>Not everyone will successfully make this transition. Just as manufacturing workers couldn&#39;t all &quot;learn to code,&quot; not all engineers will adapt to AI-centric workflows. But the overall demand for engineering talent — particularly engineers who can effectively use AI — will likely increase as the addressable market for projects expands.</p>
<p>The key insight: <strong>AI doesn&#39;t replace engineers; it changes what engineering means.</strong> Those who adapt will find more opportunities than ever before.</p>]]></content:encoded>
    </item>
    <item>
      <title>How we use AI in EPC workflows</title>
      <link>https://puranwater.com/how-to-leverage-ai-in-epc-workflows</link>
      <guid isPermaLink="true">https://puranwater.com/how-to-leverage-ai-in-epc-workflows</guid>
      <description>Not hype, table stakes: deploying deterministic, auditable AI workflows in engineering</description>
      <pubDate>Sun, 08 Jun 2025 00:00:00 GMT</pubDate>
      <content:encoded><![CDATA[<p>Ever since ChatGPT launched in research preview, Puran Water has been <a href="https://www.puranwater.com/#mcp-agents">experimenting with LLMs</a> and trying to use AI in engineering workflows. Initially the use cases were limited to research-related tasks. Helpful, but not transformative. The real breakthrough came with custom tool calling, which opened up a seemingly limitless universe of applications for any workflow that can be done sitting at a desk.</p>
<p>In EPC workflows, the most critical requirements for any automation or streamlining effort are that processes must be (1) deterministic and (2) auditable. Out-of-the-box Large Language Models (LLMs) fail to meet these criteria. Their responses vary with each prompt, and their reasoning is opaque. However, with tool-calling LLMs and properly structured workflows, achieving deterministic and auditable engineering automation is now within reach for domain experts.</p>
<h2>A tested approach to AI-assisted engineering</h2>
<h3>1. Start with software-engineering AI agents</h3>
<p>Start with SWE (Software Engineering) AI agents. One option is Anthropic&#39;s Claude Max plan, which offers incredible value for the cost with Claude Code. These agents can help you rapidly develop the infrastructure needed for your engineering workflows.</p>
<h3>2. Wrap repetitive tasks as MCP tools</h3>
<p>For every task performed more than a few times — running an engineering template, generating a datasheet, doing standard calculations — write an MCP (Model Context Protocol) wrapper that exposes it as a tool to LLMs. There are several ways to expose tools to LLMs; we build ours as MCP servers because the protocol is now cross-vendor. One server is built and maintained in one place, and any MCP client can call it: <a href="https://www.anthropic.com/news/model-context-protocol">Anthropic introduced MCP</a> as an open standard, and it has since picked up support from <a href="https://platform.openai.com/docs/guides/tools-remote-mcp">OpenAI</a>, <a href="https://www.microsoft.com/en-us/microsoft-copilot/blog/copilot-studio/introducing-model-context-protocol-mcp-in-copilot-studio-simplified-integration-with-ai-apps-and-agents/">Microsoft</a>, and <a href="https://blog.google/technology/google-deepmind/google-gemini-updates-io-2025/">Google</a>. The pattern has several advantages:</p>
<ul>
<li><strong>For engineering templates</strong>: Developing a script to replace legacy Excel or MathCAD templates becomes extremely easy, even for non-programmers who can still read and understand code with SWE AI agent assistance</li>
<li><strong>For engineering documentation</strong>: A Papermill-Jupyter Notebook approach can parameterize documents (similar to MathCAD) while making them easily accessible by LLMs via the MCP protocol</li>
</ul>
<h3>3. Build agent-driven workflows</h3>
<p>For workflows performed repeatedly, implement an agent framework (like <a href="https://www.agno.ai/">Agno</a>) to automate them with agents calling the MCP servers. Human involvement shrinks to initiating the workflow and reviewing outputs. An alternative is to parameterize Jupyter notebooks that encapsulate the full workflow (or logical chunks of it) and expose those notebooks as tools via MCP.</p>
<h3>4. Avoid over-engineering</h3>
<p>Resist the temptation to start ambitious projects. I made the mistake of attempting to create MCP servers flexible enough to perform 90% of common fluids, heat transfer, or water chemistry calculations a process engineer in my field would typically do. The open-sourced MCP servers below are a result of this effort, but the approach was flawed.</p>
<p>It&#39;s far better to build incrementally, in parallel with your current workflow, so that:</p>
<ol>
<li>You can rigorously test the tool-calling LLM output against your current tools to ensure proper functioning</li>
<li>You don&#39;t over-build tools that never get used in actual workflows</li>
</ol>
<h2>Puran Labs — MCP server demonstrations</h2>
<p>The <a href="https://www.puranwater.com/#mcp-agents">MCP server demonstrations</a> use tool-calling LLMs through a chat interface. While this format illustrates the core functionality, it represents the <strong>least powerful</strong> implementation of these systems. Chat interfaces inherently constrain LLMs to a single-threaded, human-paced interaction—essentially reducing them to interactive assistants.</p>
<p>The true power of tool-calling LLMs emerges when deployed in automated workflows that can:</p>
<ul>
<li>Execute multiple instances in parallel without human bottlenecks</li>
<li>Process batch operations at scale</li>
<li>Chain complex multi-step calculations automatically</li>
<li>Generate auditable outputs for both human and AI review</li>
</ul>
<p>Despite these limitations, the chat interface effectively demonstrates key advantages over standard LLMs:</p>
<ul>
<li><strong>Deterministic results:</strong> Each prompt produces exactly the same output (test this by running the same prompt multiple times)</li>
<li><strong>Auditable results:</strong> Click tool call badges to inspect the exact parameters passed and results returned</li>
</ul>
<p>When combined with the Papermill-Jupyter Notebook approach, these tool-calling agents create a complete audit trail that can be verified by both human engineers and AI systems.</p>
<h3>Deterministic vs. probabilistic results comparison</h3>
<p>To illustrate the critical difference between tool-calling and standard LLMs, consider this engineering calculation:</p>
<p><strong>Prompt:</strong> Size a pump to move 750 GPM of wastewater from a wet well 8 feet below grade to a treatment tank 25 feet above grade. The system uses 8-inch Schedule 40 steel pipe with 150 feet of suction line and 300 feet of discharge line. The system includes the following fittings:</p>
<p>Suction Line Fittings:</p>
<ul>
<li>1 sharp entrance at wet well</li>
<li>2 × 90° elbows  </li>
<li>1 gate valve</li>
</ul>
<p>Discharge Line Fittings:</p>
<ul>
<li>4 × 90° elbows</li>
<li>2 × 45° elbows</li>
<li>1 swing check valve</li>
<li>1 gate valve</li>
<li>1 normal exit at treatment tank</li>
</ul>
<p>Calculate the pump&#39;s total dynamic head (TDH) and NPSH available.</p>
<table>
<thead>
<tr>
<th><strong>Tool-Calling LLM (Claude Sonnet 4 with Fluids MCP Server)</strong></th>
<th><strong>Standard LLM (Claude Sonnet 4 without tools)</strong></th>
</tr>
</thead>
<tbody><tr>
<td><strong>Run 1:</strong> TDH = 38.3 feet, NPSHA = 23.6 feet<br/><strong>Run 2:</strong> TDH = 38.3 feet, NPSHA = 23.6 feet<br/><strong>Run 3:</strong> TDH = 38.3 feet, NPSHA = 23.6 feet</td>
<td><strong>Run 1:</strong> TDH = 40.0 feet, NPSHA = 23.3 feet<br/><strong>Run 2:</strong> TDH = 41.0 feet, NPSHA = 22.4 feet<br/><strong>Run 3:</strong> TDH = 51.6 feet, NPSHA = 19.5 feet</td>
</tr>
</tbody></table>
<p><strong>Analysis:</strong> The tool-calling LLM produces identical results across all runs because it uses deterministic engineering calculations through the MCP server. Each calculation follows the same methodology, uses identical fluid properties from CoolProp, and applies consistent friction factors. The standard LLM, despite being highly capable, produces different results on each run due to its probabilistic nature: variations in assumed values, calculation methods, and rounding approaches lead to substantially different engineering outcomes.</p>
<h2>Practical implementation tips</h2>
<ul>
<li><p><strong>Start small.</strong> Begin with a single, well-defined workflow that you do regularly. Build an MCP server for just that workflow, test it against your existing methods, and only then expand.</p>
</li>
<li><p><strong>Focus on auditability.</strong> Every MCP server should produce outputs that clearly show:</p>
<ul>
<li>Input parameters used</li>
<li>Calculation methodology</li>
<li>Intermediate results</li>
<li>Final outputs with appropriate units</li>
<li>References to standards or methods employed</li>
</ul>
</li>
<li><p><strong>Enable parallel workflows.</strong> The real power emerges when you can run multiple instances of a workflow simultaneously. Design your MCP servers to be stateless and thread-safe.</p>
</li>
<li><p><strong>Integrate with existing tools.</strong> Don&#39;t try to replace everything at once. MCP servers can wrap existing tools (Excel templates, Python scripts, commercial software APIs) to make them accessible to AI agents.</p>
</li>
</ul>
<h2>The path forward</h2>
<p>Tool-calling LLMs combined with MCP servers (a convenient way to give any MCP client access to custom tools) are a compelling opportunity to reimagine how engineering work gets done. We&#39;re moving from a world where engineers adapt to software interfaces to one where AI agents call custom tools on our behalf in autonomous workflows.</p>
<p>An engineer who can effectively orchestrate these agents with domain-specific tools can accomplish what previously required entire teams with the associated friction of less than perfect communication and too frequent meetings.</p>
<p>The key is to start with your specific workflows, build incrementally, and always maintain the deterministic, auditable nature that engineering demands.</p>
<hr>
<p><em>The MCP servers discussed in this article are open-sourced and available at <a href="https://github.com/puran-water">github.com/puran-water</a>.</em></p>]]></content:encoded>
    </item>
  </channel>
</rss>