← Back to home

Research Index · Updated September 2026

Active research areas

A fully expanded index of current research interests across autonomous systems, trustworthy AI, compute infrastructure, multimodal intelligence, financial AI, and machine-readable real-estate systems. Emerging areas include source references for further reading.

Research & Policy Contributions

University of Oxford

The Oxford Consultation on Copyright and AI — 2026

Named Contributor

Contributing an industry perspective to the evolving relationship between artificial intelligence, intellectual property, transparency and responsible AI governance.

View Oxford Publication →

Publications & research outputs

ACL Anthology — 2025

Contextual Grounding in Retrieval-Augmented Generation Systems

An investigation into how retrieval pipelines can be conditioned on document-level context to reduce hallucination rates and improve factual precision in long-form generation tasks. Proposes a hybrid grounding architecture combining dense passage retrieval with structured knowledge injection at the attention layer.

NeurIPS Workshop on Agentic AI — 2024

Provenance-Aware Multi-Agent Orchestration: Toward Auditable Agentic Systems

Introduces a protocol for tracking decision lineage across multi-agent pipelines. Explores how trace graphs and attribution metadata can be embedded at the action level, enabling post-hoc audit and accountability in autonomous systems deployed in regulated environments.

Research taxonomy

MCP & Model Context Protocol

  • Tool-use primitives and protocol design for context-aware LLMs
  • MCP server architectures and sandboxed execution environments
  • Dynamic context injection vs. static prompt engineering trade-offs
  • Cross-model context portability and schema standardisation
  • Stateful vs. stateless MCP session management for long-horizon tasks
  • Resource exposure patterns in MCP: files, databases, and live APIs
  • MCP as middleware for enterprise AI integration layers
  • Security boundaries and permission scoping in MCP deployments
  • Multi-tenant MCP server design for SaaS platforms
  • MCP orchestration across heterogeneous LLM backends

Agentic Search & Retrieval

  • Query planning and decomposition in agentic retrieval pipelines
  • Adaptive re-ranking with real-time feedback loops
  • Hybrid dense + sparse retrieval topologies at scale
  • Search-augmented reasoning and iterative grounding
  • Self-correcting retrieval with agent-driven query reformulation
  • Retrieval-augmented generation (RAG) for domain-specific corpora
  • Contextual compression and relevance distillation in RAG pipelines
  • Multi-hop question answering across heterogeneous knowledge sources
  • Agentic web search integration with structured output validation
  • Real-time retrieval with streaming inference and progressive summarisation
  • Evaluation harnesses for retrieval precision, recall, and faithfulness
  • Knowledge graph-augmented retrieval for structured domain reasoning

Emerging AI Topologies

  • Mixture-of-experts (MoE) routing for heterogeneous model ensembles
  • Hierarchical agent graphs and structured delegation primitives
  • Swarm intelligence patterns applied to LLM orchestration
  • Federated inference and edge-deployed reasoning networks
  • Speculative decoding and draft model acceleration strategies
  • Sparse activation networks for compute-efficient inference
  • Model merging and task vector arithmetic for capability transfer
  • Continual learning architectures without catastrophic forgetting
  • Liquid neural networks and adaptive computation graphs
  • Neuromorphic computing interfaces for energy-efficient AI
  • Mamba and state space models as transformer alternatives
  • Test-time compute scaling and inference-time reasoning budgets

Multimodal Retrieval Systems

  • Cross-modal embedding alignment (text ↔ image ↔ structured data)
  • Late fusion vs. early fusion architectures for retrieval
  • Spatially-aware retrieval for visual commerce applications
  • Captioning pipelines as retrieval preprocessing layers
  • CLIP-family models and contrastive multimodal pretraining
  • Video-text retrieval and temporal grounding in video corpora
  • Audio-visual co-embedding for multimodal document understanding
  • Zero-shot cross-modal retrieval with compositional query handling
  • Dense captioning and region-level visual feature indexing
  • Multimodal reranking with cross-attention fusion layers
  • Structured table and chart understanding in document AI pipelines
  • Unified multimodal embedding spaces for marketplace search

Diffusion Model Control & Visual AI

  • ControlNet conditioning and adapter-based fine-tuning
  • Semantic layout control for product and property visualisation
  • Inpainting pipelines for virtual product placement (VPP)
  • Consistency models and accelerated sampling for production latency
  • IP-Adapter and identity-preserving image generation
  • Textual inversion and DreamBooth for personalised visual generation
  • Latent diffusion models for high-resolution image synthesis
  • Image editing via score distillation and attention manipulation
  • Stable Diffusion XL and architectural refinements for quality
  • Controlable video generation and temporal consistency
  • Style transfer and aesthetic conditioning for brand-consistent AI imagery
  • Synthetic data generation pipelines for training vision models

Trust, Provenance & AI Governance

  • Audit trails and decision traceability in autonomous systems
  • Watermarking and content attribution for generative AI outputs
  • Regulatory compliance frameworks for high-stakes AI deployment
  • Epistemic confidence scoring and uncertainty propagation
  • EU AI Act compliance architecture for enterprise AI systems
  • Model cards, datasheets, and AI transparency documentation standards
  • Bias detection and fairness auditing in production ML pipelines
  • Red-teaming methodologies for agentic AI safety evaluation
  • Data lineage tracking from training corpus to model output
  • Constitutional AI and value alignment in deployed systems
  • GDPR-compliant AI system design and data minimisation strategies
  • AI liability frameworks and insurance risk modeling for autonomous agents

Financial AI & Behavioral Modeling

  • Irregular income pattern recognition for affordability modeling
  • Open banking data normalisation and feature engineering
  • Explainable credit and risk scoring systems (XAI for fintech)
  • Temporal behavioral trajectories for financial readiness prediction
  • Transaction categorisation and merchant embedding models
  • Fraud detection with graph neural networks and temporal anomaly detection
  • Synthetic financial data generation for model training and testing
  • Cash flow forecasting with probabilistic sequence models
  • Alternative credit scoring from behavioral and psychometric signals
  • Real-time affordability APIs for mortgage and rental platforms
  • Open finance standards: PSD2, CDR, and emerging global frameworks
  • AI-driven financial coaching and nudge system design

Real Estate Data Infrastructure

  • MLS schema fragmentation and standardisation approaches
  • Enrichment pipelines for property listing data at scale
  • Semantic property search beyond keyword-based retrieval
  • AI-driven comparative market analysis (CMA) systems
  • RESO data dictionary evolution and adoption patterns
  • Property graph modeling for neighbourhood and amenity relationships
  • Computer vision for property condition assessment and grading
  • Automated valuation models (AVMs) with uncertainty quantification
  • Geospatial embeddings and location intelligence for real estate AI
  • Document AI for lease extraction and contract intelligence
  • Real-time listing freshness detection and duplication resolution
  • Cross-border MLS interoperability and international data standards

LLM Fine-Tuning & Alignment

  • Supervised fine-tuning (SFT) strategies for domain adaptation
  • RLHF and RLAIF pipelines for preference-aligned generation
  • Direct preference optimisation (DPO) as a scalable RLHF alternative
  • LoRA, QLoRA, and parameter-efficient fine-tuning (PEFT) methods
  • Instruction tuning and task generalisation in foundation models
  • Constitutional AI and scalable oversight for value alignment
  • Catastrophic forgetting mitigation in sequential fine-tuning
  • Model merging techniques for multi-task capability synthesis
  • Synthetic data generation for fine-tuning data-scarce domains
  • Evaluation protocols for fine-tuned model safety and capability regression

AI Infrastructure & MLOps

  • Model serving architectures: vLLM, TGI, Triton, and Ray Serve
  • GPU memory optimisation: quantisation, KV cache management, paging
  • CI/CD pipelines for ML model versioning and safe deployment
  • Feature stores and real-time feature serving for production ML
  • Observability and drift detection in deployed model systems
  • Multi-cloud AI infrastructure and vendor-neutral serving layers
  • Cost optimisation strategies for high-throughput LLM inference
  • Experiment tracking and reproducibility in ML research workflows
  • Distributed training with FSDP, DeepSpeed, and Megatron-LM
  • Chaos engineering and resilience testing for AI-dependent systems
  • Model compression: pruning, distillation, and architecture search
  • Green AI and carbon-aware scheduling for sustainable ML operations

Agentic Systems & Multi-Agent Design

  • ReAct, Reflexion, and chain-of-thought agent architectures
  • Tool-augmented agents: code execution, browser use, and API calling
  • Long-horizon task planning with hierarchical goal decomposition
  • Memory systems for agents: episodic, semantic, and procedural stores
  • Multi-agent communication protocols and shared world models
  • Agentic safety: interruption, human-in-the-loop, and abort mechanisms
  • Evaluation benchmarks for agentic task completion and reliability
  • Agent identity, persona consistency, and role-based delegation
  • Simulation environments for agent training and red-teaming
  • Provenance tracking across multi-agent decision pipelines
  • Agentic workflow orchestration with LangGraph, AutoGen, and CrewAI
  • Enterprise agentic deployment patterns: guardrails, logging, and SLAs

Epistemic AI & Knowledge Representation

  • Belief revision and non-monotonic reasoning in AI systems
  • Ontology design and knowledge graph construction for domain AI
  • Epistemic uncertainty quantification in probabilistic ML models
  • Frame problem and context persistence in long-running agents
  • Causal reasoning and counterfactual inference in decision systems
  • Semantic web technologies and linked data for AI knowledge bases
  • Vector-symbolic architectures and neuro-symbolic integration
  • Commonsense reasoning benchmarks and compositional generalisation
  • Argumentation frameworks for AI-assisted decision support
  • Temporal knowledge representation and event-driven world models

Agent Interoperability & Discovery

Open protocols, capability discovery, registries, and semantic contracts for agents operating across organizational and platform boundaries.

  • MCP and A2A protocol convergence
  • Agent cards and machine-readable capability discovery
  • Cross-agent task delegation and lifecycle management
  • Semantic contracts for tools, resources, and agents
  • Interoperable identity and authorization across agent networks
  • Conformance testing for protocol implementations

Digital Provenance & Machine Identity

Verifiable origin, delegation, behavioral evidence, and policy enforcement for autonomous software and machine-initiated actions.

  • Cryptographic identity for autonomous agents
  • Delegation chains from human principal to machine actor
  • Runtime behavioral provenance and evidence graphs
  • Policy enforcement at agent and tool boundaries
  • Software and model supply-chain attestation
  • Origin verification for machine-generated actions

Confidential & Cryptographic AI

Privacy-preserving computation for sensitive AI workloads using confidential computing, encrypted computation, and distributed trust.

  • Trusted execution environments for private inference
  • Remote attestation of AI workloads
  • Fully homomorphic encryption for model evaluation
  • Secure multiparty computation across data owners
  • Privacy-preserving federated learning
  • Output controls and provenance for confidential agents

Physical AI & Embodied Systems

Multimodal perception, world models, control, and safety for intelligent systems acting in physical environments.

  • World models for planning under uncertainty
  • Vision-language-action models for robotics
  • Edge-cloud coordination for embodied intelligence
  • Simulation-to-reality transfer and synthetic environments
  • Deterministic safety layers around learned control
  • Physical-action provenance and intervention systems

Quantum Assurance & Post-Quantum Systems

Evidence, reproducibility, and crypto-agility for hybrid quantum-classical computation and quantum-resistant agent infrastructure.

  • Post-quantum migration for agent identity and delegation
  • Hybrid classical and post-quantum authentication
  • Vendor-neutral quantum execution evidence
  • Reproducibility across quantum hardware ecosystems
  • Cryptographic provenance for hybrid workloads
  • Quantum workload, hardware, and baseline characterization

Energy-Aware AI Systems

Compute orchestration that treats energy, cost, carbon, latency, data location, and hardware capability as joint scheduling constraints.

  • Energy-aware routing across heterogeneous accelerators
  • Carbon-aware training and inference scheduling
  • Performance-per-watt evaluation for model architectures
  • Edge versus cloud placement under energy constraints
  • Power-aware agent and LLM control planes
  • Evidence and reporting for sustainable AI operations

AI-Native Software Engineering

Software architectures, tests, interfaces, and evidence systems designed for humans and autonomous development agents to work together.

  • Machine-readable architecture decision records
  • Agent-safe codebase boundaries and ownership metadata
  • Executable policy in test and delivery pipelines
  • Semantic API descriptions for machine actors
  • Provenance for agent-authored software changes
  • Evaluation of AI-native development reliability

Digital Twins & Spatial Intelligence

Continuously updated spatial models linking visual evidence, sensor state, simulation, and operational decisions across physical assets.

  • Multimodal updates to building and infrastructure twins
  • Computer vision for spatial condition mapping
  • Simulation of maintenance, retrofit, and occupancy scenarios
  • Identity and interoperability for built-environment assets
  • Uncertainty and provenance in inferred physical state
  • Access control for shared operational twins