PhD Researcher, Dalhousie University · AI Lead
My research explores how language models can be adapted to perform reliably and safely in specialized technical domains.
Build high-quality technical corpora with clear provenance and coverage.
Help models learn specialized concepts, language, and task behaviour.
Connect model responses to trusted technical and regulatory sources.
Determine where models succeed, fail, or require greater caution.
Research motivation
The challenge is not simply whether a language model can produce an answer. It is whether that answer is accurate, supported, appropriately cautious, and useful to a domain expert.
Nuclear and energy systems provide a demanding environment for this research. The knowledge is technical, the language is precise, and unsupported conclusions can carry real consequences.
The broader goal is to advance language models that can operate responsibly across engineering, scientific, and regulated environments.
Research areas
The work brings together model adaptation, evaluation, evidence grounding, and data development—the foundations needed to move from general language capability to dependable specialized performance.
Exploring how compact, open-weight language models can develop stronger domain understanding and useful task behaviour without losing the capabilities that make them broadly effective.
Continued pre-training · Supervised fine-tuning · Preference optimization · Private deploymentDesigning benchmarks that reveal what a model knows, how well it uses evidence, when it becomes uncertain, and where specialization creates meaningful gains or trade-offs.
Question-only · Supplied evidence · RAG · Failure analysis · Capability retentionStudying how retrieval, agentic search, and governed tools can produce responses that are more current, traceable, and defensible for technical knowledge work.
RAG evaluation · Agentic retrieval · Function calling · Governed data accessDeveloping training and evaluation datasets from scarce technical material while maintaining the provenance, quality, and coverage required for credible research.
Question generation · Quality review · Gold answers · Dataset provenanceSelected research and publications
Each project starts with a practical limitation in specialized-domain AI, develops a research approach around it, and produces evidence that guides the next stage of the work.
More evidence and larger models do not automatically produce stronger domain understanding or more reliable regulatory reasoning.
Build a benchmark spanning CANDU engineering and Canadian regulatory material, then compare general and adapted open-weight models under different evidence conditions.
Clarifies what model scale, evidence access, continued pre-training, and supervised fine-tuning each contribute to specialized-domain performance.
Manuscript in preparation · Dalhousie University
Sensitive technical environments cannot always depend on externally hosted, general-purpose models.
Train a compact language model on public CANDU material to examine the feasibility and limitations of locally controlled domain adaptation.
Establishes a practical foundation for private domain models while identifying the data, training, evaluation, and deployment challenges that remain.
Specialized-domain answers can appear credible even when important technical details are missing or unsupported.
Create a curated technical question-answer dataset and compare direct model responses with retrieval-augmented responses through human and automated evaluation.
Demonstrates why reliable domain AI depends on the full information system—not only the underlying language model.
Static retrieval struggles as document collections grow and questions require evidence from multiple sections, sources, or levels of detail.
Trace the evolution of an enterprise retrieval pipeline toward planned and iterative information gathering, alongside governed function calling for operational data.
Shows how retrieval research can improve traceability, maintainability, and confidence in production AI systems.
Engineering and safety teams must review large volumes of unstructured operational records containing specialized terminology and inconsistent descriptions.
Develop and evaluate LLM-based workflows for safety screening, equipment-reliability events, and failure-mode identification.
Demonstrates how LLMs can support expert review while preserving accountability and clear boundaries around automation.
Research, collaboration, and applied work
I am open to research scientist and research engineer opportunities, academic and industry collaboration, and applied research contracts focused on reliable language models and evidence-grounded AI.
