Inspiration
Anyone who has used a 3D printer knows the frustration of checking on a print and realizing something went wrong an hour ago. The wasted filament is annoying, but the lost time can be worse—especially when other jobs are waiting. Our team also has an actual business selling 3D prints in our neighborhood, so we wanted an application that integrated AI for this exact purpose.
That got us thinking about small print farms. Knowing that a machine is behaving strangely is useful, but the next question is practical: What should happen to the orders waiting behind it? We built ForgeSight to connect machine monitoring with that decision.
What it does
ForgeSight gives users an interactive dashboard for a six-printer production floor. It analyzes temperature, vibration, motor current, and extrusion flow using a trained machine-learning classifier.
Users can inject simulated faults, inspect the signals behind an anomaly score, and see how those conditions affect a production schedule. The planner compares a baseline schedule with an alternative that considers deadlines, machine availability, material compatibility, and machine condition.
Users can also take a machine offline, adjust cost assumptions, import telemetry from a CSV, and export their analysis. Everything runs in the browser without paid APIs.
How we built it
We built the interface with HTML, CSS, and JavaScript. For the machine-learning component, we used Python and scikit-learn to train a four-class logistic regression model on synthetic sensor windows.
The model recognizes normal operation, extrusion blockage, thermal instability, and mechanical vibration. We exported its learned weights so classification could run directly in the browser.
For scheduling, we implemented deadline-aware dispatch followed by local search over machine assignments and job order. The planner keeps the baseline whenever its proposed changes fail to improve the calculated objective.
We used ChatGPT/Codex to assist with implementation, testing, and documentation.
Challenges we ran into
Connecting the model to a useful business decision required more thought than simply displaying an anomaly score. We needed explicit assumptions about disruption, material waste, and late orders so users could understand where the estimates came from.
Scheduling also introduced constraints that were easy to overlook. A printer cannot handle two jobs simultaneously, and moving a PETG order to a PLA-only machine does not solve anything.
Another challenge was keeping our claims realistic. A model that performs well on generated data still needs testing on actual equipment. We made that distinction visible throughout the application.
Accomplishments that we're proud of
We completed a working path from sensor data to classification to a proposed production schedule.
The model was trained on 3,600 synthetic windows and evaluated on 1,200 held-out windows. We also included computational tests covering classification, schedule feasibility, cost calculations, and CSV validation.
Our favorite feature is the ability to change a scenario and immediately inspect the consequences. Users can see the evidence, adjust the assumptions, and decide whether the proposed plan makes sense.
What we learned
A machine-learning score becomes much more useful when it connects to a decision someone actually needs to make.
We also learned why evaluation needs context. Strong results on synthetic data demonstrate that the pipeline works within the simulator; they do not establish reliability on real printers.
Making those limits understandable became part of the product design.
What's next for ForgeSight - Intelligence & Predictive Maintenance Platform
Our next step is collecting labeled telemetry from real printers and evaluating the model across different machines and operating conditions.
From there, we want to calibrate the scores, add camera-based defect detection, and compare scheduling recommendations with actual production outcomes. That would let us move toward predictive maintenance backed by real equipment data.
Log in or sign up for Devpost to join the conversation.