Jim Galvan

Each game object combines its Three.js mesh with its Rapier physics body, keeping the visual and physical representations together. Cars accelerate through mass- and time-based impulses, while steering uses torque to produce movement that feels physical rather than scripted.

The Rust server uses Tokio to manage concurrent connections and distribute each player's position, rotation, and tire state. It also acts as the shared source of truth for player names, colors, and starting positions.

The inference pipeline prepares images through contrast normalization, edge enhancement, sharpening, and noise reduction. This produces more consistent visual input across different lighting conditions, camera angles, and image quality levels.

I also created repeatable retraining workflows that use targeted data augmentation to incorporate new image scenarios and production failure cases. Each model update is evaluated against previously learned examples to improve detection without introducing regressions.

The application combines Grounding DINO and SAM to locate food-related regions and generate detailed segmentation masks. Instead of treating a meal as one isolated object, the pipeline can associate related elements such as the plate, toppings, sides, and surrounding ingredients.

A weighted scoring algorithm evaluates the detected regions and selects the most likely main subject when several dishes or food items appear in the same image.

The inference environment supports GPU acceleration with PyTorch and can also run on a Raspberry Pi for edge deployments. The application is packaged with Docker and served through a Django backend.

Status: Private repository; application in development

Businesses can organize menu categories, add item descriptions and photos, customize their branding, generate QR codes, and share mobile-friendly menu pages with customers. The platform also supports printable menus, localized content, and reusable business information across the customer-facing experience.

The backend is built with Java, Quarkus, and PostgreSQL, while the frontend uses Angular and TypeScript. Uploaded images are processed into multiple sizes and stored in AWS S3 so each page can load an appropriate resolution instead of the original full-size file.

The platform also includes Spanish localization, email verification, account onboarding, payment-related workflows, and production error monitoring. Sentry provides visibility into failures and unusual user states that may not appear during local testing.