FastAPI ML Model Management Platform

This FastAPI application provides a platform for managing and interacting with YOLO machine learning models for computer vision tasks. Built during a paid ML training course (~30 students learning computer vision through hands-on project work), I volunteered to build the REST API and deployment infrastructure as an additional contribution, leveraging my prior full-stack development experience while most students focused on data annotation and model training tasks.

The platform features session-based model selection, video processing at 2-second intervals with frame extraction and GIF generation, and shared media management. Successfully deployed on Render as a demo for the course's client project—detecting objects in satellite imagery from Russia/Eastern Europe regions. The models were trained on this specific geographical landscape and perform well for that context, though performance may vary for other regions. Users can test it using screenshots from Yandex Maps or Google Maps satellite views of similar terrain.

**Key Learning Project**: This was my first ML deployment project, built rapidly to demonstrate the trained models for the course deliverable. While functional for the demo use case, it has several architectural shortcomings that reflect my growth since then:

• **Architecture**: The entire application lives in a 700+ line main.py file with global state for model management. Today I'd structure this with proper separation of concerns (routes, services, models) and use dependency injection instead of global variables.

• **Security**: No authentication, no input validation on file uploads, potential path traversal vulnerabilities, and unsafe model loading without torch.load(weights_only=True). A production system would need JWT/API keys, rate limiting, and proper validation.

• **Code Quality**: Uses print() statements instead of proper logging, has commented-out code blocks, and inconsistent type hints. The deprecated PIL.textsize() method needs updating to textbbox().

• **API Design**: Lacks versioning, pagination, and standardized error responses. The /project_structure endpoint is an information disclosure risk that should be removed.

Despite these issues, the project successfully delivered a working demo to the client and taught me core ML deployment concepts: handling video processing pipelines, managing PyTorch memory in production, implementing async file uploads, and deploying containerized ML services. Taking initiative to build this API infrastructure while learning alongside the ML team taught me both technical skills and the value of leveraging existing experience to contribute in new domains.