Inspiration
Our inspiration was to move beyond reactive dashboards. In high-stakes environments like HP Metal Jet production, knowing what happened is not enough; engineers need to know what might happen. We were inspired by the concept of Monte Carlo simulations applied to hardware reliability, creating a bridge between raw physics formulas and real-world uncertainty.
What it does
Chronotwin is an AI-powered Digital Co-Pilot designed to transform how operators interact with the HP Metal Jet S100. It moves beyond static monitoring by combining a physics-based simulation engine with probabilistic forecasting.
The system operates in two core modes:
Deterministic Simulation: Models real-time component wear (e.g., Thermal Resistors, Recoater Blades) based on direct input drivers like temperature, humidity, and metallic dust contamination.
Probabilistic Forecasting: Runs hundreds of parallel pseudo-aleatory simulations to generate a reliability distribution, predicting the likelihood of failure before it occurs.
Agentic Diagnostics: A natural language interface where an AI Engineer analyzes telemetry, cites evidence from drivers, and provides immediate, 24h, and preventive action paths via voice or text.
How we built it
We developed a modular stack focused on "Physics meets AI":
Simulation Engine: Built using Python, we implemented mathematical formulas that define health decay as a function of environmental and human factors.
The Brain: Integrated Gemini 3.1 Lite as the reasoning core. We built a custom tool-calling layer (llm_tools.py) that allows the model to query historical data and run diagnostic functions.
Frontend: A high-performance Streamlit dashboard featuring custom CSS, "Health Bar" components, and interactive Plotly charts for driver correlation.
Challenges we ran into
The "Hallucination" Trap: Early versions of the AI would guess why a part failed. We had to implement a strict Grounding Protocol that prevents the agent from responding unless it can cite specific telemetry evidence (e.g., "Contamination exceeded 0.3 units").
Latency in Reasoning: Running complex agentic loops with multiple tool calls was initially slow. We optimized this by moving to the Gemini 3.1 architecture and streamlining the tool definitions.
State Management: Synchronizing a live simulation engine with a voice-enabled chat interface in Streamlit required advanced session-state handling to prevent UI resets during AI generation.
Accomplishments that we're proud of
Dual-Mode Engine: Successfully bridging the gap between a replay simulation and a stochastic simulation that calculates real failure probabilities.
Zero-Hallucination Diagnostics: Achieving a system that strictly follows HP maintenance protocols, providing actionable engineering advice instead of generic LLM chatter.
Voice Integration: Implementing a hands-free "Co-Pilot" mode where engineers can ask for status updates and hear the diagnosis while working on the machine.
What we learned
We learned that Industrial AI requires more than just a large language model, it requires Domain Grounding. By encoding the physics of the HP Metal Jet S100 into our simulation formulas, we provided the LLM with a "physical intuition" it wouldn't have otherwise. We also gained deep experience in Agentic Workflows, learning how to structure tool-calling to make an AI act as a proactive partner rather than a reactive search engine.
What's next for ChronoTwin
Fleet-Wide Sync: Expanding the digital twin to monitor multiple S100 units simultaneously, identifying cross-machine patterns (e.g., if a specific batch of powder is causing wear across the factory).
Automated Parts Ordering: Integrating the "Preventive Action Path" directly with supply chain APIs to automatically queue replacement parts when health drops below 20%.
Edge Deployment: Moving the simulation engine closer to the hardware for sub-millisecond response times to critical thermal anomalies.
Built With
- gemini-3.1-lite(api)
- github
- javascript
- json
- python
- streamlit
Log in or sign up for Devpost to join the conversation.