I build systems that turn intent into action. The work splits into two understandings, and they only connect if both are described in the same terms: the person, and the options, with the actions attached.
The first is reading the person: intent inference, the signal under the request, the goal behind the click. The second is the one that usually goes unbuilt, and the one I work on at Velou: understanding the products themselves. A shopper’s intent is only as useful as the data it lands on, and ecommerce product catalogs arrive underdescribed: a title, a handful of fields, whatever a merchandiser had time to type. I described each individual product into one deep, normalized ontology, millions of them, so an intent could find the right item instead of the nearest keyword. You can’t connect dots you don’t understand.
With both sides deep and in one language, the rest is connection: rank intent to the right option and act. Then measure what actually happened and feed it back: the harnesses, golden datasets, regression suites, and calibrated judges that tell you a change helped before you trust it further. Two rules hold under all of it. A connection is only as good as the weaker of the two understandings. And it only gets better if you close the loop.
The options so far have been products and suppliers. The machine doesn’t care: dinner tonight, a trip, a calendar, any set of options that can be described connects the same way.
I’m the founder of Velou, and this is the work; some of it is patented.