Research
AI SYSTEMSRESEARCH.
I'm interested in the space between what AI systems are built to do and what they actually do when deployed. That gap - in behaviour, in reliability, in safety - is where I focus.
RESEARCH INTERESTS.
Active Areas of Inquiry01
Action Boundaries in Agentic Systems
How should autonomous AI agents define, enforce, and communicate the limits of what they will and won't do? I'm interested in the specification, detection, and graceful failure of boundary violations - and what "graceful" actually means when real tasks are at stake.
AI Safety · Agent Design02
Evaluation of AI Systems in Production
Benchmarks measure performance in controlled settings. Production measures something entirely different. I'm exploring evaluation frameworks that capture reliability, consistency, and real-world failure modes - not just accuracy on test sets.
Evaluation · Reliability03
Retrieval-Augmented Generation Architecture
RAG is now a standard pattern, but its failure modes are poorly understood. I'm interested in chunking strategy, retrieval quality, context injection, and how these interact with model behaviour - particularly in low-resource and domain-specific settings.
RAG · Architecture04
Technical AI Safety - Applied
Not AGI alignment theory - practical safety for the AI systems being deployed today. Adversarial inputs, prompt injection, data leakage in RAG systems, and how to build pipelines that fail safely and auditably rather than silently and catastrophically.
AI Security · SafetyRESEARCH
PHILOSOPHY
PHILOSOPHY
"I'm not interested in models that perform well on benchmarks. I'm interested in systems that behave reliably when someone depends on them at 2am for something that actually matters."
PUBLICATIONS.
Peer-Reviewed WorkApplication of Mathematical Convolution Approach of Image Sharpening to Digital and Satellite Imaging
Applies mathematical convolution theory to practical image sharpening problems in digital photography and satellite imaging. Demonstrates how pure mathematical abstractions translate directly into computational image processing pipelines.
Read Paper →ONGOING WORK.
In ProgressWIP
Action Boundary Specification for LLM Agents
A framework for formally specifying, testing, and enforcing action boundaries in LLM-powered agents - with a focus on practical implementation rather than theoretical proofs.
In Progress · 2026WIP
RAG Failure Mode Taxonomy
A structured taxonomy of failure modes in production RAG systems - from retrieval failures to context injection artifacts - with diagnostic patterns and mitigation strategies.
In Progress · 2026