CASE 1/4 DH Premium: a subscription where nobody believed one could exist
Selling a subscription to a free-news audience in price-sensitive markets was widely assumed impossible. The assumption was an average. Averages lie.
Before building anything, I went looking for who might pay. The decisive finding: users consuming 3× median content showed 5× the lifetime value as subscribers. Willingness to pay wasn't spread thin across 300M users; it was concentrated in a heavy-reader core. That one correlation shaped everything after it: the product was built for the core, and the funnel was built to find them.
The moves
- Launched DH Premium 0→1: ad-free reading, premium content, early access.
- Pricing found by experiment, not benchmark: price-point tests across cohorts before committing.
- Lifecycle over broadcast: a tiered re-engagement system keyed to engagement depth. Each tier got different messaging, offers, and timing.
The win wasn't the subscription. It was refusing to average: 300M users don't have one willingness to pay. Find the segment the economics work for, then build the machine that finds them at scale.
CASE 2/4 Short News: a format bet that had to prove it wasn't cannibalism
Attention was migrating to short-form everywhere. The open question wasn't "can we build it," it was whether a snackable format would grow engagement or just eat the feed. I led the 0→1 launch: format definition, content pipeline, feed integration, and instrumentation designed specifically to separate net-new engagement from cannibalization.
CASE 3/4 AI content-risk detection: trust as a product surface
At 300M MAU in a charged news environment, one viral piece of misinformation is a business risk, and manual review doesn't scale. With ML engineers I built content-risk detection using narrative classification and publisher-credibility signals.
My contribution was the part a model can't decide for itself: defining what "good" means. I owned the eval metrics, built the feedback loops that kept the model honest, and set the precision/recall trade-off from the product side, because over-blocking legitimate publishers is its own harm.
CASE 4/4 The unglamorous layer: instrumentation
I owned event taxonomy and funnel tracking end-to-end. That analytical plumbing is what made everything above measurable, and it directly cut onboarding drop-off by 3%. Years later, the same discipline is why I catch measurement artifacts other teams read as product problems.