AI research + software

Robust AI.
Useful apps.

I research medical visual question answering and build software for people, institutions, and public services.

Conceptual medical AI pipeline turning a noisy radiograph representation into clean visual tokens
Noise-aware representation learning, visualized.

Making medical AI less brittle.

A denoising autoencoder learns to recover clean visual embeddings before they are projected into a language model.

Noise-Aware Visual Representation Learning for Medical Visual Question Answering

I Putu Adi Pratama, Bahadorreza Ofoghi, Atul Sajjanhar, Shang Gao

The framework improves robustness to noisy visual embeddings while maintaining competitive performance on clean inputs.

Corrupt

Train with controlled noise in visual embeddings.

Reconstruct

Recover clean features with a denoising autoencoder.

Project

Map visual features into the language model space.

Answer

Adapt efficiently with LoRA for medical VQA.

Selected applications

Software made for real contexts.

Academic infrastructure, publishing, exam practice, and public data exploration.

Research should hold up under noise. Software should hold up in use.

I work across machine learning and software engineering, moving from careful experiments to interfaces that make complex systems approachable.

Medical AI Visual question answering and robust representations
Applied machine learning Efficient adaptation, evaluation, and reliability
Public-facing systems Clear tools for education, research, and civic information

Start with the paper.

Read the complete method, experiments, and findings on arXiv.

Read the research