Leon Shi

Systems and studies

Leon Shi

Decision Scientist & AI Engineer

I’m drawn to complex systems and to the ways incisive decisions can compound into outsized outcomes.

My work spans resource allocation, synthetic experimentation, derivatives and crypto markets, and scientific discovery.

I’m based in NYC, where outside of work you can find me playing water polo, LARPing as an interior designer, or tracking down a café with a good mocha.

Pencil portrait of Leon Shi smiling
Portrait
Featured drift

MSM01.001

Product resource allocation

At Audible, I own the analysis behind a recurring review of product investment with the Chief Product & Analytics Officer.

MSM01.001 · Internal analysis · In recurring use

Product investment review

I own the analysis behind a recurring review of product investment at Audible. It puts product areas on comparable terms, connects earlier experiments to current choices, and gives the next cycle something concrete to revisit.

Question
Where should product investment go next?
Method
Bring product costs, returns, and experiment evidence into a comparable view.
Use
A recurring review with Audible’s Chief Product & Analytics Officer.

MSM02.014

Synthetic experimentation

I built a synthetic cohort to test participant-level predictions against experiments that have already run.

MSM02.014 · Internal method · Evaluated prototype

Synthetic experiment reads

The system simulates a varied cohort, aggregates its responses into an experiment-level read, and compares that read with historical web experiments. The evaluation rewards agreement and error; fluency is irrelevant.

Cohort
A varied set of simulated participants, built for the experiment context.
Evaluation
Historical web experiments are the reference for agreement and error.
Status
Internal method under evaluation.

MSM03.027

Cronos

A live research product for replaying options strategies against historical data.

MSM03.027 · Independent product · Live

Cronos

I built Cronos end to end: market data, strategy definition, contract resolution, portfolio replay, research interface, authentication, billing, and deployment. A user can define a strategy, resolve its contracts, and inspect the resulting path.

Open Cronos
Research path
Strategy → contract selection → historical replay.
Built
Data pipelines, research kernel, interface, accounts, billing, and deployment.
Status
Live.

MSM04.063

Physics search

I use language models to widen the candidate set, then let numerical checks decide what survives.

MSM04.063 · Research prototype · Ongoing

Scientific discovery prototype

The prototype searches topological-material and dark-matter parameter spaces. A model proposes candidates; numerical evaluators test them against explicit constraints. Rejected candidates remain part of the search record.

Search
Generate candidates across a defined parameter space.
Gate
Run numerical checks before a candidate enters review.
Status
Research prototype in active development.

MSM05.089

Research notes

Questions and working methods that sit behind the projects.

MSM05.089-A · Working essay · Open

Post-AI society

A working essay on what changes when capable machine work becomes cheap, concentrated, and widely available. It follows the consequences for ownership, status, attention, and the institutions that decide who gets access to what.

Read the working essay

MSM05.089-B · Working paper · Open

Endogenous knowledge diffusion in frontier AI

This paper models an industry where deployment has two effects: it lowers the leader’s cost through experience and gives competitors a way to catch up. The model combines limit pricing, experience curves, and an endogenous diffusion rate. It predicts persistent concentration with leadership churn.

Read the working paper
Mechanism
Leader deployment improves its own efficiency and increases the rate at which rivals learn.
Prediction
Concentration can persist while the leading firm changes.
Status
Working draft, July 2026.

MSM05.089-C · Technical note · Live

How I generated the drifts

A technical note on the shared WebGL renderer behind the lab, museum, and hyper review: one point field, shader-generated drifts, scroll transitions, pointer input, and deliberate curation.

Read the technical note