Research agenda

I build reliable AI systems for clinical, speech, multimodal, agricultural, and scientific applications.

My research asks how learned representations can remain useful, accountable, and resilient when they meet people, sensitive data, and changing environments.

Healthcare & Clinical AI

Clinical AI work focused on privacy-preserving language systems, uncertainty, unlearning, and dependable decision support.

This domain treats privacy and reliability as model-design constraints rather than post-processing concerns.

Current questions

  • How can clinical language systems protect sensitive tokens without discarding useful context?
  • How can uncertainty and unlearning improve the reliability of multimodal clinical models?

Methods

  • Client-side token selection and cryptographic redaction
  • Vision-language modelling, uncertainty estimation, and selective unlearning

Privacy, Trust & Safety

Trustworthy AI research spanning encrypted inference, secure information handling, and robust detection of harmful inputs.

The emphasis is on explicit threat, utility, and evidence boundaries for applied systems.

Current questions

  • What privacy utility trade-offs appear when only sensitive content is protected?
  • How can evaluation expose brittle behaviour under constrained or adversarial inputs?

Methods

  • Cryptographic protection and privacy-utility evaluation
  • Robustness testing across curated and external datasets

Speech, Audio & Synthetic Media

Speech-forensics work on cross-paradigm detection, generator attribution, open-set recognition, and emotional manipulation.

This line of work connects geometric representation learning with practical forensic questions about provenance and generalization.

Current questions

  • Which representation geometries remain useful across synthesis paradigms and languages?
  • How should a detector respond when a generator was not present during training?

Methods

  • Speech foundation models with hyperbolic, spherical, and Euclidean fusion
  • Graph-based attribution and confidence-aware open-set inference

Multimodal & Human-Centered AI

Multimodal learning for human traits and performance, with datasets and models designed around naturalistic evidence.

The work begins with the quality and structure of human-centred data, then carries those constraints into model design and evaluation.

Current questions

  • How can video, audio, and language representations retain complementary structure?
  • How should model capacity and geometry be balanced for practical training?

Methods

  • Multimodal dataset construction and pairwise comparative annotation
  • Manifold-specific experts with adaptive routing and tangent-space fusion

Scientific & Applied AI

Scientific machine learning for thermal-property prediction using augmentation, ensembles, and metaheuristic search.

Scientific applications provide a test of whether a model remains useful beyond benchmark settings and under limited observations.

Current questions

  • How can scarce experimental data support high-fidelity property prediction?
  • Which combinations of augmentation, ensembles, and optimization are useful across thermophysical tasks?

Methods

  • Literature-curated experimental datasets and controlled augmentation
  • Ensemble learning, neural models, and metaheuristic hyperparameter search

Earth & Agricultural Intelligence

Computer vision and decision support for soil, rock, crop, and medicinal-herb intelligence.

This domain links visual evidence to practical decisions in agriculture and earth science while keeping dataset scope explicit.

Current questions

  • How can visual models handle small, heterogeneous datasets in environmental settings?
  • How can predictions become actionable recommendations without hiding uncertainty?

Methods

  • Transfer learning, CNN-ViT ensembles, and image augmentation
  • Fuzzy decision systems and cross-validation for applied recommendations