This page is focused on the engineering layer after model training: exporting models, building APIs, connecting databases and object storage, designing frontends, containerizing services, and deploying AI applications to real cloud environments.
My apps are not just demos. They are practice grounds for model serving, application structure, persistence, automation, and deployment decisions.
Build custom models or adapt pretrained architectures using PyTorch, TensorFlow, Keras, and Hugging Face.
Move from raw .pt or .keras artifacts toward portable .onnx inference assets.
Expose inference through FastAPI and Uvicorn with validation, routing, prediction metadata, and app-level logic.
Save predictions, metadata, logs, and uploaded image references using PostgreSQL, SQLite, MSSQL, or object storage.
Containerize with Docker and launch to Render, Hugging Face Spaces, or similar platforms.
Only two apps right now, but this page is designed to scale as more trained models become product-grade interfaces.
End-to-end medical imaging application: classification ensemble, segmentation model, FastAPI backend, Docker deployment, PostgreSQL prediction storage, Supabase image storage, Hugging Face model hosting, and MLOps layers with MLflow, Airflow, and local Kafka simulation.
Trained disaster tweet classifier served as a web application with FastAPI, SQLite logging, Jinja frontend templates, Docker packaging, and a clean prediction interface for real-time text inference.
Besides AI deployment work, I also build applications in other computer domains. These projects focus more on software engineering, desktop interfaces, data structures, and system design than on model inference.
A desktop educational management system built outside the AI domain: a C++ and Qt Widgets application for handling national records, students, teachers, courses, terms, grades, reports, and user access in one structured interface with file-based persistence and clean desktop workflows.
A compact map of the technologies I use to convert models into real interfaces, services, storage layers, and cloud-ready systems.
Exporting models is only one layer. The bigger focus of this page is how trained systems become real applications through API serving, database integration, cloud storage, interface design, and MLOps orchestration.
My main deployment interest is exposing predictions through structured endpoints with request validation, preprocessing logic, model inference, and response formatting.
I care about connecting applications to databases for prediction history, metadata, uploaded file references, audit trails, and application-side analytics.
This part covers storing models and uploaded assets, then connecting them to actual serving targets such as Render, Hugging Face Spaces, and model registries.
Beyond the model, I want to practice building clear interfaces for uploads, forms, prediction cards, status feedback, and portfolio-grade presentation.
My end goal is not only deployment, but also tracking, orchestration, and structured workflow control around the application stack.
I still care about moving from raw checkpoints toward portable assets like .onnx, especially when deployment constraints make lighter inference formats more practical.
