Damian Tran

Graph neural entity resolution

At Minerva I built foundational AI systems and end-to-end data pipelines from the ground up to pull signal out of open-source intelligence: identifying criminal and bad-actor behaviour in a mountain of open-world data.

The approach I developed for consolidating records across sources is graph neural entity resolution — connecting the dots between structured and unstructured data much the way a human investigator does. Around it I built data-curation applications that automate acquisition, labelling and quality control with bias reduction, feeding a feature-engineering approach in high-dimensional embedding space for edge-annotation networks to learn within.

The result was an investigation tool with complex entity-resolution capability and high sentiment and risk-classification accuracy, used by Fortune 500 companies on millions of real users in production. Heavily regulated enterprises cut review time for alerts and investigation cases from hours, sometimes days, to minutes.

• Graph Neural Networks
• Entity Resolution
• Multimodal Data Integration
• Financial Compliance AI

The geometry of learned representations

I read latent spaces through the lens of differential geometry, asking what the geometric properties of a learned representation tell us about how a model behaves and where it generalises. Curvature, geodesics and the local metric are not metaphors here; they are the objects that decide which points a model treats as near.

I am drawn to differential geometry for the same reason I am drawn to Tadashi Tokieda’s work on the visuospatial basis for mathematics. A short form can carry a genuinely complex idea about the world, rooted in nothing more than changing your view and your reference frame across spaces of different dimension.

• Differential Geometry
• Representation Learning
• Deep Learning

Neurosymbolic reasoning with a consistent world view

My own experimentation now applies graph representations of multimodal data to neurosymbolic reasoning, aiming at models whose world view stays consistent. The interest is in frameworks that integrate symbolic reasoning with neural architectures through graph-based representations, so a system can hold a coherent model of the world across modalities rather than one per input type.

This is the thread that connects the work: entity resolution is a symbolic claim about identity made from statistical evidence, and getting it right at scale requires both halves to be first-class.

• Neurosymbolic AI
• Agentic AI Systems
• Real-time AI Architectures
• NLP