Teaching Hindsight to Detect Project Risks

"How did it know we were falling behind on the backend?" I questioned, only to find Hindsight had been silently learning from our task completion velocity to calculate a project risk score that was uncomfortably accurate. What I Built and Why I wanted a project manager that could actually observe how a team works over time and reason about those patterns—not just display a Kanban board with coloured chips. Specifically: could I build something that watches task data accumulate over weeks and eventually surfaces something like "Charlie's documentation sprint is at risk" based on nothing more than historical completion patterns and deadline proximity? The result is an AI Project Manager: a FastAPI backend (main.py) backed by a SQLite database (managed through SQLAlchemy in database.py and db_models.py), with a React + Vite frontend. The AI layer is handled by two components working in tandem—Groq as the LLM for language generation, and Hindsight for persistent agent memory. The whole thi