About
Twenty years on one question: how decisions get made when the evidence is incomplete. First in brains, now in systems.
Now
I am the AI and Agentic AI Chapter Lead at Commonwealth Bank, where I lead the AI for Customer Service Excellence chapter — AI scientists and developers delivering agentic AI, generative AI and machine learning across wealth and private banking and the customer service network. Earlier I led the AI Innovation chapter in the Chief Data and Analytics Office, working across technology, retail banking, customer advocacy and fraud prevention.
I work across the full life cycle — problem shaping, solution design and architecture, through to production and governance. Recent work has been customer-facing agentic systems for investment education and trading guidance, built on retrieval-augmented generation with the safety and risk evaluation framework a regulated environment requires, alongside agent-to-agent orchestration for customer validation, embedding-based scam detection, and large-scale entity resolution.
A regulated bank is an unusually good place to build AI, because it forces the question most teams postpone: how do you know this is working, and how would you know if it stopped? Most of what I do is evaluation, governance and getting things into production, not modelling.
Before
I trained as a cognitive neuroscientist. My PhD, at Western Sydney University with the RIKEN Brain Science Institute, used a Stroop-based paradigm and MEG to study attentional capture by emotional faces and hand gestures.
Eight years of postdoctoral work followed, at the University of Sydney, RIKEN in Japan and Stanford. At Sydney I used transcranial magnetic stimulation to interfere with visual cortex, and found that weak stimulation — below the threshold at which people report seeing anything —improves detection of faint stimuli rather than degrading it. I also developed REPT, a faster method for estimating phosphene thresholds, since replicated by independent labs.
The thread that mattered most began at RIKEN and finished at Stanford, with collaborators at UCL: how a person's recent history of choices biases their next decision. People do not weigh each new piece of evidence afresh. They carry the last few trials with them, and they treat failure differently from success. That work was published in PNAS in 2016 and is still the research I return to most.
Why the two halves are the same job
Psychophysics is the discipline of measuring a noisy decision system that cannot tell you how it works. You present controlled evidence, record the choices, and infer the mechanism from the pattern of errors. Signal detection theory separates how well a system can discriminate from how willing it is to say yes — and that distinction, which every vision scientist learns in their first year, is the one most often missed when people evaluate a model by its accuracy.
I came to enterprise AI with that as a habit rather than a framework, and it turns out to be the useful part. Knowing where a system goes wrong under uncertainty is worth more than knowing how to build it.
Also
Before the bank I was the first data science hire at an education technology startup, building a production classifier over ten million websites. Before any of it, a Master's in applied mathematics and computer science at Yerevan State University.