A predictive maintenance and Digital Co-Pilot system for the HP Metal Jet S100 3D printer.

Links: Devpost · GitHub

Overview

Blue Lobster combines a synthetic digital twin, reinforcement learning maintenance planning, telemetry dashboards, and an LLM-backed diagnostic assistant to serve as an industrial operator workspace.

System Architecture

The project is structured into three integrated layers to handle complex machine telemetry and maintenance reasoning:

• Physics and Machine Learning: Instead of relying on a physical machine, we built a digital twin using PyTorch that simulates the degradation of critical components. A Deep Q-Network (DQN) agent learns maintenance strategies by balancing intervention costs against failure risks.
• Digital Co-Pilot Backend: A FastAPI application bridges the database and the AI. It uses Google Gemini as a diagnostic agent, strictly grounded to explicit database ticks. Time-series telemetry is compressed using pytoony to optimize token usage.
• Interactive 3D Frontend: A Next.js App Router and Tailwind CSS dashboard provides the operator interface. It features a 3D digital twin built with React Three Fiber, allowing users to scrub through time, inspect hardware failures, and interact with the floating AI co-pilot.

Tech Stack

• Frontend: Next.js 16, Tailwind CSS, React Three Fiber, Three.js
• Backend: FastAPI, SQLite, Google Gemini API
• Machine Learning: PyTorch, Deep Q-Network, tabular Q-learning