I work on calibrated uncertainty: when a model says 85%, does 85% happen?
The Exceedance Design Effect · code and Lean proofs
A dataset does not have one effective sample size. In our document experiment, the same 1,000 rows carried about 217 independent observations' worth of information at the median, and about 621 at the 95th percentile. The number of rows is a property of the dataset. The effective sample size belongs to the analysis.
SureStop · pip install surestop
Stop bad training runs early, with a conformal bound on how often you stop a good one. Measured on 400 LoRA fine-tuning runs under a pre-registered protocol. Ships as an Optuna pruner in OptunaHub.
Mondrian Decomp 1 · series · weights
75 paintings, each a p5.js sketch written by an open model and chosen by hand from thousands of generations. Every sketch re-renders its painting pixel for pixel.
Soothsayer
You pick the coverage level, and the oracle returns the price band that empirically delivers it. Validated on 12 years of weekend gaps across 10 tickers.
- scryer · Rust data platform that writes versioned, deduplicated Parquet for market research.
- solana-dlmm-meteora · Pure-Rust swap math for Meteora's DLMM, with no RPC.
- restate-elixir · Elixir SDK for Restate. Passes all 202 official conformance tests.
Director of Data & AI at Digital Locke · adamnoonan.xyz · ORCID · LinkedIn




