Ten years across data science and product, three products, and the studies that decided what stayed in them.
Products
Each one exists because something kept breaking in the ordinary way of doing it. Each carries a pre-registered study of a question that mattered to its design.
A news pipeline that assigns each extracted claim a verification tier, keeping the strength of the evidence separate from how well the claim was reported, so a well-reported claim cannot be upgraded to a true one.
23.9%
of week-one Contested verdicts held up against a later verdict built from cited evidence, so about 3 in 4 of the claims marked Contested were true.
Local discovery for people who are past the touristy layer, from a layered retrieval engine that reads intent from a plain question and is built to surface the places list sites bury.
69.4%
of good answers lost when query interpretation is removed, the largest cost of any single layer, an estimate from fewer runs than the headline figure.
Synthetic user populations that browse a real web app and turn what they do into findings: rage clicks, confusion, pages nobody can find.
64/64
Every session that produced the behavior its rage-click signal looks for was caught, and none was caught without it, in a study that maps where the instrument goes blind.
Studies
Designed before the data existed, run against the products I build, and ending in a recorded decision. Every number cites the committed code or data it comes from.
Audit
VerifiedNews
Week-one Contested verdicts were re-judged against later cited evidence, and about 3 in 4 of those claims turned out true. The fix to the single-denial rule is decided and queued behind work that has a consumer for the tiers.
Teardown
Sidestreet
A 9-arm ablation priced every layer of the engine. The strict filter stayed off: restoring it costs 16.7% of good answers to prevent a quality loss the study could not find.
Five seeded defects at controlled doses, rerun under resampled populations, drew a boundary that extends trust to element-level detections and withholds it from page-level signals on their own and from silence in retention mode.
Substack, 2026-08-07
Notes from building B2B lookalike modeling without a warehouse: the architecture, the constraints, and why novelty comes from data layers.
Every study has its own repository, with its frozen design, its data, its analysis, and a script that recomputes published numbers from source. The studies and their code open with each release.
Frozen design, 600-claim sample, re-scoring pipeline and verifier.
9 arms, stability gates, and the cost of every retrieval layer.
Seeded defects, synthetic populations, detection boundaries.
This site: content as typed collections, and gates that fail bad builds.
$ python3 specs/verify_published.py → 30 checks OK, exit 0
About
I've spent ten years where data science meets product: enterprise data at Autodesk, then six years at 6sense working with enterprise accounts, where a value-measurement framework I built outside my job description became a product adopted across teams. Now I'm the data scientist at Pristine Data AI, heading data product and strategy. I built the data function there from nothing, and the analysis sits inside the product decisions it feeds.
The products on this page started outside work and turned into real systems. The studies exist because I wanted my own decisions held to the standard I'd demand at work.
Ask this site anything: a retrieval engine over everything here is in design. For now, the studies answer the hard questions.