What I keep coming back to is the gap between how things look on paper and how they actually work: how people behave, and how a decision really gets made. That's the thread running from behavioral science to climate finance. The material changed; the habit of watching closely didn't. Mostly I've learned to trust a good question more than a tidy answer.
The course used redacted documents from a real wind project. My individual final model connected generation, operating cash flow, debt constraints, and equity returns. It could show how a structure behaves under stress, but it was not my transaction and it could not make the decision for me.
That is the useful boundary of a model: it lays out the judgment so its assumptions and consequences can be inspected.
Continue reading →A four-month technoeconomic brief on where quantum technologies could create material value in energy, and on what timeline. Our six-person team compared grid optimization, nuclear research, battery chemistry, enhanced solar, energy finance, and quantum sensing through a shared three-axis framework.
Coursework built around a redacted real-project case: contract reviews and an individual wind model covering operating cash flow, PTCs, debt sculpting, P50/P99 cases, and scenario analysis.
Teaching support for Lisa Sachs's graduate course on the architecture of climate finance: blended and concessional capital, debt sustainability, carbon markets, the UNFCCC framework, and cost-of-capital drivers in emerging markets.
In January, quantum computing was a vocabulary problem. By April, it had become a judgment problem.
You can run ten thousand sensitivities and still have to make the call yourself.
The structure of a deal says who absorbs the first loss, and on whose word the rest of the capital shows up.
A vertical AI agent for U.S. renewable energy project finance. It parses PPAs, validates sponsor models, checks IRA-era tax equity structures, and drafts IC memos with citations. Phase 0: eval harness on hand-graded public PPA cases, not a product yet.
An AI rewrites the bullets in a LaTeX résumé to match a specific job description, without making things up. I built it during the SIPA on-cycle market when every tailoring round took an hour by hand. Pairs with Claude Code (via `claude -p`) or any LiteLLM-compatible model.
A notes-first workspace for turning meetings, messy notes, and audio into structured bilingual documents. It pairs a Markdown editor with transcription, project memory, AI synthesis, and export paths for PDF, Markdown, and Feishu.
I trained as a behavioral scientist. Four years of facing a messy human situation and asking what exactly is going on here, and what we mean by "going on".
That same question turned out to be right for project finance, too. A pro-forma is a disguised argument about the future, and the assumptions are the footnotes.