I shipped two production Claude agents and a 0-to-1 risk model inside a regulated insurer, each gated on an evaluation built first, with uncertainty methodology reviewed by the financial-market regulator.
I develop the evaluation before the product.
#5
Founding technical team member, CarbonPool
2
Production AI agents shipped, gated on evaluation
~97%
Faster analytical turnaround, rebuilt with Claude Code
$XX M
Pre-underwritten premiums in year one, behind the risk model
IPCC AR6
Contributing Author · 24 papers · 4,000+ citations
Career Arc
Four chapters, each one adding a capability, converging on AI product leadership in regulated, high-stakes domains.
Chapter 1
As an IPCC AR6 Contributing Author I quantified uncertainty for the Paris Agreement, work that is part of 24 peer-reviewed papers and 4,000+ citations in journals including Nature Climate Change and Nature Geoscience. I won the CHF 760K SNF Ambizione grant as sole investigator, then declined it to build commercial product instead.
Capability addedMeasuring what a model does and does not know.
Chapter 2
Primary specialist on TCFD physical-risk disclosures for enterprise clients, translating asset-level climate analysis into decisions a C-suite would act on.
Capability addedTurning technical analysis into executive product under a regulated frame.
Chapter 3
Co-developed the Scope 3 methodology behind a Net-Zero Planner SaaS, built the data pipelines, and designed the platform storylines serving enterprise clients.
Capability addedShipping a data product end to end.
Chapter 4
Founding technical team member (#5). Built the 0-to-1 risk model behind a first-of-kind carbon-credit insurance product, then shipped two production AI agents, all under insurance liability.
Capability addedBuilding and evaluating AI where a person signs the output.
The thread is one decision repeated:
develop the evaluation first, because the signature is mine.
Guiding Principles
Evaluate before deploying, and keep high-stakes decisions understandable and contestable.
“…social justice is not only a goal to be safeguarded after technologies are deployed, but a condition that must shape their very design from the outset.”
by Pope Leo, Magnifica humanitas, 109
“…when data and algorithms influence credit distribution, personnel selection or access to services and opportunities, it is necessary that decisions be understandable, contestable and subject to oversight, so that individuals are not reduced to mere profiles.”
by Pope Leo, Magnifica humanitas, 164
About
I build AI and data products for high-stakes, regulated domains: insurance capital, carbon credits, counterparty due diligence. The common thread is liability. A person signs the model's output, so I develop the evaluation before the product.
I am the founding technical team member (#5) at CarbonPool, where I led the product and data strategy for the risk-modelling stack behind the world's first in-kind carbon-credit insurance product, and built and shipped production LLM agents with Claude Code. The decade of applied research behind me is the reason the rigour holds: ten years measuring uncertainty for the IPCC, then three years shipping AI under insurance deadlines, and the habit survives the commercial pressure.
I can brief a board member, an underwriter, and an engineer in the same afternoon and lose none of them. That is the job.