Most of the 1,300 tools in this repository are calculators: a bid, a fee, a tax threshold. On 2026-10-20 the owner set a new direction: tools built from methods in published research papers, run on the visitor's own data, licences respected and the papers credited, and usable by AI agents as well as people. The rules are in portfolio/research-tools/README.md. This post covers the first three. AI coding agents (Claude sessions) built them, with the owner steering.
Each one runs in a browser tab with nothing uploaded. Each ships a zero-dependency CLI (node cli.mjs --help prints the citations) and an engine that the repository's agent layer can call (node platform/agents/tools.mjs describe <slug> now prints each paper in full). Each comes with a real public dataset as its example.
1. When did this series change? Changepoint detector
Two methods run side by side:
- PELT (Killick, Fearnhead & Eckley 2012, JASA) finds the exact best segmentation under a penalty per change.
- Bayesian online changepoint detection (Adams & MacKay 2007) gives a probability for each candidate change.
The example is the annual flow of the Nile at Aswan, 1871–1970. That is the textbook case: Cobb (1978) found the level dropped after 1898. The tool reports the same thing:
- segments 1871–1898 (mean 1097.75) and 1899–1970 (mean 849.97);
- a Bayesian probability of 79% at 1899, the only point above 50%.
That matches R's changepoint package (cpts = 28).
Two tests check the implementations are exact, not approximately right:
- PELT's answer equals a brute-force search over all segmentations on 25 random series.
- With the hazard driven to zero, the Bayesian recursion's evidence equals the closed-form Normal-Inverse-Gamma marginal likelihood.
One honest number from the tests: on pure noise of length 300, the default BIC penalty still invents a change in about one series in ten.
2. How wide must a prediction interval be? Conformal prediction intervals
Conformal prediction gives intervals that cover a chosen share of new cases, with a finite-sample guarantee and no assumption about the error distribution. The tool fits three variants and checks each on held-out rows:
- split conformal (Lei et al. 2018);
- a locally weighted version;
- conformalized quantile regression (Romano, Patterson & Candès 2019), with its own exact linear-programming quantile fit.
On the diabetes data from Efron et al. (2004), at a 90% target with one seeded split, split conformal covered 90.1% of the 111 test rows (100 of 111) at a mean width of 182.0. CQR covered 93.7% at 180.7. Averaged over 200 random splits, all three land between 90.2% and 90.4%. That is what the guarantee promises: correct on average over splits, noisy on any one split.
3. What repeats, and what is unlike anything else? Matrix profile explorer
The matrix profile (Yeh et al. 2016; Zhu et al. 2016, both ICDM) gives, for every window of a series, the distance to its closest match elsewhere in the series. Low values are motifs (repeated patterns); high values are discords (anomalies). The tool computes it exactly with STOMP. It agrees with a brute-force implementation and with the outputs of STUMPY, a widely used library, to within 1e-10 on the test series.
On NOAA's Mauna Loa monthly CO2 record, with a 12-month window:
- The closest repeat is the year starting September 2014 and the year starting September 2024 (distance 0.1394; correlation 0.9992).
- The most unusual year is the one starting January 2024 (distance 0.7474).
A discord says a stretch is unusual, not why. The README notes that NOAA's file includes months measured at Maunakea rather than Mauna Loa near then.
What's next
The spec lists the next papers, each one a job someone has:
- synthetic control, for "did my intervention work";
- Kaplan–Meier and Cox models, for churn;
- Sinkhorn optimal transport;
- Gaussian-process regression;
- isolation forests;
- robust PCA;
- STL decomposition;
- false-discovery-rate control.
Desktop packaging is also queued. The repository already builds iOS, Android, npm, MCP and browser-extension packages, and these tools qualify for them.
These are statistical tools on data you supply. They are not advice. Check any result before you rely on it.