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.

It can't make up a number, because every figure is checked against the statistics engine I wrote from scratch. Tukey runs it in the cloud. Gosset is my own model, trained on about 15,000 examples from that engine, and runs offline on your own graphics chip.

This deviceChecking your hardware.

Trend following backtest

• Python
• pandas
• SciPy
• bootstrap
• unit tests

A backtest of the 50/200 day moving average crossover on SPY, written from scratch with next day trades, trading costs and unit tests. I tuned it on 2000 to 2017 and tested on 2018 to 2026.

It didn't beat buy and hold. Then I checked whether picking the windows was just luck. Only 3 of 89 pairs beat buy and hold, and the probability of backtest overfitting came out around 0.57, basically a coin flip.

A/B testing library, checked by simulation

• Python
• SciPy
• pytest
• CLI

Real experiments break the textbook t test. People check results every day, metrics are noisy, splits come out uneven and dashboards track twenty metrics at once. I built a Python library that handles each of those, with CUPED and always valid sequential tests.

I checked every method on thousands of simulated experiments, then ran it on Cookie Cats, a mobile game that split 90,189 players. It found a 4.3% drop in 7 day retention and caught that the groups were split unevenly.

What it costs to turn off the internet

• SQL
• DuckDB
• causal inference
• 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 that stayed online, using GitHub's public data on how much code every country pushes.

During Iran's 2026 blackouts its GitHub activity ran 69% below its twin. Across six full shutdowns, each day offline cost about 1.3 days of work, so people didn't catch up after. I planted 1,500 fake shutdowns first to make sure the method finds them.

Live simulator

What gets an app to 100K installs

• Python
• pandas
• scikit-learn
• 2.2M rows

I followed 815K Google Play apps from launch to see which ones ever reach 100K installs. Only about 6 in 100 do, and how an app makes money mattered most. Free apps with ads and in-app purchases broke out 31% of the time, paid ones 1.7%.

The gap held in all 48 categories, not just games, though part of it runs backwards since apps often add purchases once they have users. Then I turned the model into the simulator here.

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

Desktop widgets for Windows. The main one shows the song that's playing and recolors itself from the album art, and the chrome text is rendered with NumPy. It only redraws what changed and pauses during fullscreen games, so idle CPU is close to zero.