AeroGlass XAI
- 5 Devlogs
- 10 Total hours
A real-time flight telemetry simulator and auditable predictive safety dashboard for next-gen jet engine maintenance.
This needs changes before we can consider it shipped. Review the Shipwright feedback, make updates, then ship again.
Feedback
Your code seem to be largely written by AI. This can be seen from both the code and the user interface. As a remainder, up to 30% of AI usage is allow for Stardance challenge, and exceeding the AI usage would not be accepted. Please rewrite your code to make sure you meet the limit.
If you have any questions, feel free to ask in #ask-the-shipwrights.
Shipwright walkthrough
This needs changes before we can consider it shipped. Review the Shipwright feedback, make updates, then ship again.
Feedback
Hi,
Both your demo links cant be accepted, the one on the stardance page leads to a bad gateaway error and the one in your github leads to a stremlit application. Please resubmit with an acceptable demo.
Shipwright walkthrough
This project was vetted by Shipwrights and is now up for voting. You can ship again after making more updates.
Feedback
Nice project, I'd love to see what you make in the future!
For any questions dm @Frog on slack or or email frog@shipwrights.dev
Shipwright walkthrough
Ship #2
Changes requested
Migrated AeroGlass XAI to a Dedicated Server (Hack Club Nest)
The Nest server uses an aggressive reverse proxy that is currently rate-limiting some of the dynamic JavaScript chunks required to render the heavy Plotly/SHAP visualizations.
This needs changes before we can consider it shipped. Review the Shipwright feedback, make updates, then ship again.
Feedback
We can't accept demos hosted on Streamlit or equivalent, as they have a loading time and may not stay up. Please find an alternative provider like Hack Club Nest. DM @Carlson for questions.
This project was vetted by Shipwrights and is now up for voting. You can ship again after making more updates.
Feedback
Hey! Your project is amazing! I really enjoyed exploring all the features you’ve included. Everything is already great and I believe this site will be life-changing! (I'm not sure what these numbers mean:)) Overall, I really enjoyed your project, everything works great! You’re very talented, and I wish you the best of luck moving forward. If you have any questions, feel free to DM @barbe <3
Shipwright walkthrough
Ship #1
Changes requested
I spent this session getting AeroGlass XAI completely ready. I successfully deployed the dashboard to Streamlit Community Cloud and handled some caching optimizations for the heavy SHAP tensors using mmap_mode.
The project is officially live!
The core ML architecture is done, so my final focus was transforming AeroGlass XAI into a production-ready enterprise tool.
The ‘GO / NO-GO’ Pre-Flight Briefing: I built a new operational dashboard that allows dispatchers to select the specific port and starboard engines mounted to an airframe. The app runs a real-time check against the Attention BiLSTM. If the RUL dips below the 30-cycle threshold, the UI triggers a dynamic HTML/CSS ‘NO-GO’ lockdown, pulling the top 3 thermodynamic stressors from the SHAP array to justify the grounding to the mechanics.
Financial Interactivity (The ROI Sandbox): Predictive maintenance is only useful if it saves money. I upgraded the Operational ROI Calculator with a dynamic slider that simulates the ‘AI False Negative Rate.’ This stress-tests the financial model in real-time. If you push the AI’s error rate too high, the dynamic metrics instantly flip from Cyan (‘Millions Saved’) to Red (‘Millions Lost’)
The finish line is near the corner. It will be ready to ship soon!!!
The core engine of AeroGlass XAI is now fully operational. The goal of this project wasn’t just to predict jet engine failure but to force the neural network to explain its reasoning to human engineers.
Over the last couple of hours, I successfully wired the backend prediction arrays into a dynamic, cycle by cycle Explainable AI (XAI) interface using Streamlit and Matplotlib.
The Breakthrough: I caught the BiLSTM doing something fascinating while I was building the visualizer. Near the end of an engine’s life, the model was aggressively penalizing the Remaining Useful Life (RUL) based on Core Speed (S9) but simultaneously extending it based on HPC Static Pressure (S11). Because speed and pressure are physically correlated (Multicollinearity), the AI was mathematically ‘balancing’ its weights.
By wrapping the SHAP values in a color-coded UI (Red for Core Thermodynamic, Blue for Mechanical), the dashboard immediately exposed this AI logic flaw to the human eye. This is exactly why safety critical aerospace requires ‘Glass Box’ interpretability.
I’m currently building AeroGlass XAI, a web-based command center for jet engine maintenance. While I have a robust backend foundation using the NASA CMAPSS FD004 dataset, my focus for this hackathon is the “Glass Box” transition. I am taking pre-computed BiLSTM predictions and SHAP arrays and wrapping them into a production-ready Streamlit architecture.
I have just finalized a scalable, multi-file directory structure to ensure this isn’t just a script, but an enterprise-grade application. I’ve also implemented the native dark-mode aerospace aesthetic in the configuration.
My aim is to turn static math into a “Live Telemetry” simulator and an interactive XAI audit tool where users can drill down into thermodynamic sensor impacts in real-time.
