Matin Nikoomanzar

Beta · waitlist open

Skeptic, an AI for people who care about the numbers

• JavaScript
• statistics
• AI agent
• on device model
• Python package

Most AI sounds sure of itself even when it's wrong. Skeptic checks every number first, its own included, and tells you in plain words what holds up. Show it a chart, a claim, an article or your data.

• Catches charts that lie and redraws the honest version.

• Fact checks whole articles, one claim at a time.

• Tries to break your conclusion before others do.

• Shows its work, with the Python and R code that proves it.

• Lets you change anything it makes, by hand.

Every figure is checked against the statistics engine I wrote from scratch, and anything it can't verify gets flagged. Tukey runs it in the cloud. Gosset is my own model, fine tuned on 12,000+ worked checks from that engine, and runs offline on your own graphics chip.

This deviceChecking your hardware.

Trend following as portfolio protection

• Python
• pandas
• block bootstrap
• Reality Check
• 123 tests

I tested the 50/200 day moving average rule on a century of US stocks and 35 years of markets outside the US, to see if it works as insurance. I wrote the question, the rules and what would count as a pass down first and froze them before running anything.

It lost less than buy and hold in 9 of the 10 big US bear markets since 1927, about $21,000 ahead per $100,000 at the bottom, and that cost about $5,400 a year in normal times. But just holding 69% stocks and the rest in T-bills did about as well for the same risk, and my precommitted test came out Mixed.

A/B testing library, checked by simulation

• Python
• SciPy
• sequential tests
• pytest
• CLI

Real experiments break the textbook t test. People check results every day, revenue is skewed, users have many sessions and splits come out uneven. abkit is a Python library and command line that handles those, with CUPED, always valid sequential tests, group sequential designs and a written plan before the test starts.

I tested every method with 127 checks on made up experiments where I knew the answer, many of them on messy data, and the validation table says where each one breaks. That also caught two bugs in my sequential test code. On Cookie Cats, a real test with 90,189 players, the groups came out uneven (p = 0.009), which is a warning sign worth chasing down.

When a country's internet goes dark

• SQL
• DuckDB
• synthetic control
• Python
• pytest

When a government shuts off a country's internet, there's no A/B test to run. So I built each country a synthetic twin out of 106 countries with no shutdown in my registry, using GitHub's public data on where code gets pushed from, and measured the gap. Then I spent most of the time trying to break it.

Iran was offline for about half of early 2026, and its pushes came in 59% to 69% below its twin. It's a range because most countries jumped about 30% that quarter and sanctioned ones mostly didn't, though Venezuela grew 48%. OONI's daily network tests line up with the blackout dates. It's GitHub's map, so people on VPNs counted elsewhere are part of the gap.

Live simulator

Launch-day odds for Play Store apps

• Python
• scikit-learn
• survival analysis
• 2.2M apps

Most new Google Play apps go nowhere. Of the apps about a year old in June 2021, 13% had 10K installs, 3.5% had 100K and 0.6% had 1M.

A model using only launch-day facts like category, title and the developer's history ranks 2019 launches at an AUC of 0.84. My first version said 0.86, but it used fields read in 2021, long after launch. The simulator adds intervals and real apps like yours. Money shows up there only as a pattern.

Household size prediction

• Python
• LightGBM
• R
• team of three

A class Kaggle competition, predicting how many people live in a household from their Amazon purchases. Since the answer is a small count, we treated it as classification and used the expected value of the predicted probabilities, which beat regressing on it directly.

TF-IDF on product titles beat our hand picked keywords, and with more features than rows we solved ridge in dual form. A stack of seven models placed 3rd of 23 teams.

Aura desktop widgets

• Python
• PyQt6
• NumPy
• Windows APIs
• 253 tests

Desktop widgets for Windows. A now playing card that takes its colors from the album cover, GIF boxes, and a clock in 3D chrome rendered with NumPy. Version 1.1 adds five card shapes, an analog clock face, dragging the progress bar to seek, snapping, a hide shortcut, saved profiles, a settings window with a live preview and an installer.

It also fixes a leak in the fullscreen check that kept about 270 MB a day per widget, makes GIFs play at the right speed (0.96x instead of 0.65x for 20 ms frames), and keeps every line of text at 3:1 contrast or better on 122 test covers. I've only measured it on Linux so far.