Min Ihm Data Science & Product

10 years in data science and product. Explore what I've built and how I got there.

See the products Read the research

Products

VerifiedNews

The news, read claim by claim, with each claim marked by how much backs it.

Built because a claim repeated across outlets is not the same as a claim that is true, and no reader has time to check.

See VerifiedNews

Week 40 Sep 28 to Oct 4, 2026

Economy

PCE Inflation Dips to 3.4% in August Reading

4 outlets V1×5V2×5V5×2

V4 Contested Factual 5 sources
“More than 5% of Swiss glaciers' ice has vanished in 2023 alone.”
Contested: web sources contradict claim: Swiss glaciers lost about 4% in 2023, while losses exceeded 5% in 2022, not 2023.
Sources (5)
- npr.orgC09/30
- bbc.co.ukC09/30
- washingtontimes.comR09/30
- news.google.com10/01
- news.google.com09/30
Web-contested Live web sources
Swiss glaciers lost about 4% in 2023, while losses exceeded 5% in 2022, not 2023.
apnews.com · swissinfo.ch · bbc.com · npr.org · aa.com.tr · phys.org · cnn.com · meteonews.ch · bbc.com · cbsnews.com
SRQ 0.483Bias spread 2/5Factuality 0.59Bias div 0.30

Sidestreet

What to do around a place, planned as a whole day.

Built because maps and list sites rank for everyone at once, and the places worth finding rarely make the first page.

See Sidestreet

Understudy

A product tested by thousands of synthetic users before any human sees it, with a trust level on every finding.

Built because user research is too slow and too small to run on every release, and an AI finding needs a measured trust level before anyone acts on it.

See Understudy

An Understudy finding card, a rage click on the Quick Plan control, with its recommendation

Understudy readiness: the current integration level, guest, and what its testing already produces

Research

How often do week 1 verification verdicts hold up against a later verdict built from cited evidence?

VerifiedNews, audit

VerifiedNews marks each claim in its first week and never looks back, so the question was whether those early verdicts deserve the trust the product puts in them. Rechecking 600 of them months later showed the supported tiers holding and the Contested tier failing. 23.9% stayed Contested, which changed the rule that marks a claim Contested.

What does each retrieval layer buy, and what does strictness cost a discovery engine?

Sidestreet, teardown

The engine's strict occasion filter had been removed in July, and nobody could say what it had cost or what the other layers were buying. Removing each layer in turn put a price on every one, and the filter stayed off, since restoring it lost 16.7% of good answers for a quality gain that never showed up.

Can findings from synthetic user populations be trusted?

Understudy, stress test

Synthetic testers are only useful if their findings can be trusted, so 5 known defects were planted in a real app to see which ones Understudy caught. Element-level findings held up across resampled populations, with 47.4% of sessions that hit a planted dead control caught, and page-level signals did not, which set the boundary on what the product may claim.

Memos

Aug 7, 2026 The System Is the Easy Part: Building Lookalike Modeling Under Real Constraints (leaves this site) Substack

On GitHub

The research repositories open with each release.

github.com/minihm

verifiednews-rescoring-study: Frozen design, 600-claim sample, rescoring pipeline and verifier.
sidestreet-ablation-study: 9 arms, stability gates, and the cost of every retrieval layer.
understudy-detection-study: Seeded defects, synthetic populations, detection boundaries.
portfolio-site: This site, with content as typed collections and gates that fail bad builds.

About

Define problems with customers, build products with data.

I have spent 10 years in data science and product, starting with enterprise data at Autodesk, then 6 years at 6sense with enterprise accounts, where a value measurement framework I designed outside my job description became the standard value reporting for the customer success organization. At Pristine Data AI I am the data scientist heading data product and strategy, where I designed the data function and the evaluation standard the product ships on.

I am building toward leading data and AI products where measurement is part of the design, so that whether a product works is answered by the product itself.

Autodesk

Data analyst and engineer

2016 to 2019

A graph query layer over the enterprise account hierarchy resolved most parent and child relationships automatically and cut manual review by more than 80%.

6sense

Principal data consultant

2019 to 2024

A value measurement framework built from each customer's own history became the standard value reporting for the customer success organization.

Pristine Data AI

Data Scientist — Head of Data Product & Strategy

2024 to now

The account fit and lookalike scoring the product is priced on, released on an incrementality gate.