The latest articles on DEV Community by 尤煒立 (@_01370e9a4eacc43fbc48b). https://dev.to/_01370e9a4eacc43fbc48b https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4102190%2F4c2df441-2c86-4324-928b-db12b8669d2a.png https://dev.to/_01370e9a4eacc43fbc48b en 尤煒立 Mon, 31 Aug 2026 06:10:58 +0000 https://dev.to/_01370e9a4eacc43fbc48b/why-llms-fail-at-project-scheduling-cpm-and-how-we-fixed-it-with-graph-theory-published-true-22mp https://dev.to/_01370e9a4eacc43fbc48b/why-llms-fail-at-project-scheduling-cpm-and-how-we-fixed-it-with-graph-theory-published-true-22mp <h1> Why LLMs Fail at Project Scheduling (CPM) and How We Fixed It with Graph Theory </h1> <p>If you have ever tried asking ChatGPT or Claude to generate a 120-task master construction schedule, you noticed three fatal flaws:</p> <ol> <li>Broken Dependency Loops: Tasks that depend on each other cyclically.</li> <li>Calendar Blindness: Scheduling concrete curing over national holiday shutdowns.</li> <li>Output Friction: Outputting raw text instead of native Microsoft Project (.mpp).</li> </ol> <p>Here is how we solved this problem using a decoupled AI Agent Skill and a deterministic graph engine in Python.</p> <h3> Decoupled Architecture </h3> <ul> <li>Semantic Parsing (LLM) -&gt; Deterministic CPM Engine (Kahn's Algorithm) -&gt; Native .mpp Export.</li> </ul> <h3> Try the Web Demo &amp; Samples </h3> <ul> <li>Web Demo: <a href="https://lead-pm-landing.vercel.app" rel="noopener noreferrer">https://lead-pm-landing.vercel.app</a> </li> <li>GitHub: <a href="https://github.com/kentyu0922/pmp-lead-pm-scheduler" rel="noopener noreferrer">https://github.com/kentyu0922/pmp-lead-pm-scheduler</a> </li> <li>Early Bird (50% OFF code: EARLYBIRD): <a href="https://kentyu.gumroad.com/l/pmp-lead-pm-scheduler/EARLYBIRD" rel="noopener noreferrer">https://kentyu.gumroad.com/l/pmp-lead-pm-scheduler/EARLYBIRD</a> </li> </ul> ai architecture productivity