Why LLMs Fail at Project Scheduling (CPM) and How We Fixed It with Graph Theory published: true tags: python, ai, architecture, productivity

Why LLMs Fail at Project Scheduling (CPM) and How We Fixed It with Graph Theory

If you have ever tried asking ChatGPT or Claude to generate a 120-task master construction schedule, you noticed three fatal flaws:

1. Broken Dependency Loops: Tasks that depend on each other cyclically.
2. Calendar Blindness: Scheduling concrete curing over national holiday shutdowns.
3. Output Friction: Outputting raw text instead of native Microsoft Project (.mpp).

Here is how we solved this problem using a decoupled AI Agent Skill and a deterministic graph engine in Python.

Decoupled Architecture

• Semantic Parsing (LLM) -> Deterministic CPM Engine (Kahn's Algorithm) -> Native .mpp Export.

Try the Web Demo & Samples

• Web Demo: https://lead-pm-landing.vercel.app
• GitHub: https://github.com/kentyu0922/pmp-lead-pm-scheduler
• Early Bird (50% OFF code: EARLYBIRD): https://kentyu.gumroad.com/l/pmp-lead-pm-scheduler/EARLYBIRD