top of page

Available · Q3 2026 / NYC / Product design · AI systems

I design the decisions AI products make — not just the screens.

Product designer focused on the line where AI meets real consequences. What the model is allowed to decide. What has to be a rule. How the system shows its work. Currently shipping case studies on retrieval memory and LLM policy architecture.

01 · Featured work

Two case studies about the hardest part of AI design — deciding what the model doesn't decide.

Both projects started from the same observation: LLM products fail in predictable ways. These case studies are what the design response looks like when you take that seriously.
 

2026 · in depth

02 · Selected work

Earlier work — five years across B2B, social, finance, and accessibility.

Earlier work — seven years across B2B, social, finance, and accessibility.

2020 – 2025

03 · About

Designer by training, systems thinker by reflex.

I'm a product designer based in New York. Most of what I've worked on for the last seven years lives in the boring, important space between what a product does and what the user sees — information architecture, data-heavy dashboards, permission-gated flows, the quiet structural decisions that make or break trust.

In the last year my work has moved toward AI systems: not the interface layer on top of a language model, but the design question underneath it. Where does the model belong in the product? What does it decide, and what gets decided for it? How do you show a customer that a system is telling them the truth?

I prototype aggressively. The case studies on this page are not mockups — they're built. That's where the real design decisions show up.

Based in

New York City · working globally

Focus

AI systems · product · UX architecture

Latest

The Translator Pattern · Apr 2026

Background

5 yrs · B2B · social · finance · a11y

04 · How I work

Six things I'm actually good at — not six things I'd list on a CV.

05 · Get in touch

Building something where the design decisions are actually hard?

Stress test: prompt injection · role override · emotional manipulation · policy misquote · authority claim · malformed parameter. Evaluation was synthetic (not production traffic) — real deployment would need live A/B and human-in-the-loop review sampling. The point of this pattern is that the architecture makes the right behavior cheap, not that passing six scripted attacks is proof of production-readiness.

bottom of page