Materialmind

Inspiration Manufacturing defects in raw materials cost industries billions of dollars annually. Invisible inconsistencies—micro-cracks, discoloration, foreign particles—only reveal themselves downstream, causing costly production failures, recalls, and rework. Quality control teams spend excessive time manually inspecting batches and still miss subtle anomalies. We built MaterialMind to democratize high-quality inspection using open-source AI, reducing waste and enabling faster, data-driven production decisions.

What You Learned Nemotron 3.5 Lightning: NVIDIA's open-source Mixture-of-Experts model with 30B total parameters and 3B active parameters, optimized for high-volume execution in always-on agents. Benchmarked 30% faster than Qwen3 35B on PinchBench at comparable accuracy. Convolutional Autoencoders: Reconstruction-based anomaly detection where the model learns to compress and reconstruct normal material appearances; anomalies produce high mean squared error (MSE) in reconstruction, quantified via $MSE = \frac{1}{N}\sum_{i=1}^{N}(x_i - \hat{x}_i)^2$. Streamlit Deployment: Python-native framework for rapid UI prototyping; streamlit run app.py launches a local web server with GPU-accelerated inference via workerd/Nebius Token Factory. MoE Architecture Insights: How Mixture-of-Experts models selectively activate 3B parameters out of 30B per token, enabling efficient inference without computing all 30B parameters for every request.\

How You Built It

Architecture mermaid

flowchart TD A[Upload Image] --> B[CV Autoencoder] B --> C[Anomaly Score & Heatmap] C --> D[Nemotron 3.5 Lightning Report] D --> E[Natural Language Inspection Report]

style A fill:#f9f9f9,stroke:#333,stroke-width:1px
style B fill:#e1f5fe,stroke:#015796,stroke-width:1px
style C fill:#e1f5fe,stroke:#015796,stroke-width:1px
style D fill:#f3e5f5,stroke:#7b1fa1,stroke-width:1px
style E fill:#f3e5f5,stroke:#7b1fa1,stroke-width:1px