Every AI product has two lives. The first is the demo: a hand-picked request, a strong model, an impressed room, and nobody asking what it cost. The second starts when real traffic arrives, and it runs on different fuel: cost per request, quality nobody is watching anymore, latency at the 99th percentile, a bill that grows faster than usage. Most teams prepare carefully for the first life and improvise the second.
I'm a product manager who builds the layer underneath AI products: the economics, evals, and decision infrastructure that decide whether they survive production. The full argument: After the Demo.
whisperly.io · LinkedIn · AI Economics series
Whisperly is a paid macOS dictation app I build and run solo: it turns messy speech into clean, ready-to-send text. It is also where this layer gets tested on real users. The eval harness gates every release, and unit economics is a product constraint, so every model swap is a margin decision. Its open-source predecessor is Murmur.
Every repo here started as a decision I needed a number for. Six stages, one direction:
ai-economics → ai-ab-testing-tool → multi-model-router → ml-impact-calculator → product-plugins → whisperly.io
Text version
economics → evals → routing → impact → skills → production
what it costs → is it better → which model → quality to money → repeatable work → where it survives
AI Economics: 26 posts written March to May 2026, from token pricing through model selection and agent cost patterns to unit economics. The interactive calculator implements the math. Read the series or open the calculator.
Shipping Whisperly for macOS, prototyping its iOS app, writing build-in-public essays on LinkedIn. Updated August 2026.