Making visual evidence understandable.
Exploring how vision-language understanding and digital image forensics can complement each other when analyzing insurance claim imagery.
Visual inputs
Claim images and context
Multiple signals
Semantic and forensic
Interpretation
Assess consistency
Useful output
Evidence for review
Can we look beyond surface-level similarity?
Images can be semantically consistent yet still contain suspicious modifications. Conversely, unusual pixel-level patterns do not automatically mean the story behind an image is false. This project examined how different kinds of visual evidence could be brought together to support better analysis.
The central design idea was to treat image understanding and forensic inspection as complementary perspectives rather than interchangeable solutions.
Combining different ways of seeing.
Vision-language representations
Use learned image and text embeddings to compare visual content with descriptions and related inputs.
Manipulation-related signals
Study anomaly detection and forensic techniques that look for traces less visible to semantic models.
Structured analysis pipeline
Organize complementary methods so their outputs can be interpreted together rather than as isolated predictions.
Good models still need good systems.
Part of the challenge is deciding when each analysis is useful, how to communicate uncertainty, and how to avoid treating a model score as a definitive verdict.
- Separate semantic understanding from low-level forensic evidence.
- Design for reviewable outputs rather than black-box conclusions.
- Consider operational tradeoffs when moving from experiments toward scalable pipelines.
- Evaluate methods in their actual context; benchmarks are only part of the story.
The lesson: context matters.
What I value most from the project is the experience of connecting machine learning research with engineering questions: what a signal really means, how it should be validated, and how a system can support human judgment.
This page is a high-level, public-safe overview of research themes. It does not publish employer-specific architecture, private datasets, or confidential implementation details.
Back to selected work