Trust you can verify

Security of AI systems

Large language models and the agents built on them are now exposed directly to untrusted input, which makes them an attack surface rather than just a capability. I work on defences that are cheap enough to run in the request path: dense semantic embeddings for prompt-safety guardrails, benign-anchored methods for cutting false alarms in intrusion detection, and identity mechanisms that let autonomous agents prove who they are acting for.

LLM prompt safetyguardrails agentic AISSI for AI agents explainable AIintrusion detection

Key publications

LLM and generative model architectures

Current direction — work in progress

Most of my work so far has treated large language models as something to defend. I have recently started working on building them instead: autoregressive LLM architectures, and diffusion-based generative models as an alternative to left-to-right generation. The two paradigms trade off differently on sample quality, controllability and inference cost, and those differences matter for the safety questions in my other work — a model you can steer during generation is a model you can constrain.

autoregressive modelsdiffusion models generative modellingLLM pretraining model architecture

Privacy-preserving identity

Identity verification normally works by disclosure: to prove you are eligible, you reveal the document that proves it. My doctoral research attacked that trade-off, building electronic know-your-customer systems where a user proves a claim without surrendering the underlying data — using self-sovereign identity, decentralized identifiers, verifiable credentials and zero-knowledge proofs over Merkle-tree commitments.

SSIDID verifiable credentialseKYC zk-SNARKsZK circuit design MPC

Key publications

Blockchain and secure communication

Distributed ledgers are the substrate for most of this work, so I spend time on the substrate itself: consensus behaviour, smart-contract security, post-quantum-safe design, and what it takes to anchor credentials on-chain without the cost becoming absurd. Applied deployments have covered supply-chain traceability, healthcare authentication, charitable giving and on-chain governance.

consensussmart-contract security post-quantum-safe designWeb3 DeFidApps

Key publications