Two openings

Seven products, and four of them want a model doing something that is not allowed to leave the device.

Alpha Gallery already computes perceptual hashes, a blur score and event grouping on the phone, because the promise is that no photograph is ever uploaded to be looked at. Alpha Health decodes lab reports and writes a week in review. Alpha Match plans the logistics of a date and is forbidden from writing anyone’s messages. Alpha Wealth reads receipts.

None of that is calling an API and pasting the answer back. The interesting constraint is that most of it has to run on a mid-range Android phone, inside an install budget, with no server that has ever seen the input.

What you would be doing

• Choose, shrink and ship models that run on the device — quantisation, distillation, and knowing when a classical algorithm beats a model outright
• Decide honestly what cannot run locally, and design the opt-in for the times it genuinely has to leave
• Own quality end to end: the eval set, the failure cases, and the thing people actually complain about

What we would expect

• You have shipped inference on a phone, not only in a notebook
• Python for the training side, and enough Dart, Kotlin or Swift to own the part that ships
• You can say “a model is the wrong tool here” and mean it

A bad fit if: You want to work on frontier training runs, or you want the model to be the product. Here it is a feature inside something else, and the something else matters more.

Open in three places, and where you are changes the job.

Two of the products were built for the Gulf and for India before anywhere else: Alpha Wealth and Alpha Health both ship in English, Arabic and Hindi, with right-to-left layout written in from the first commit rather than retrofitted. The company itself is registered in New Mexico. Those are three genuinely different markets with three different app stores, three different ideas of what a trustworthy finance app looks like, and very little that carries across.

What you would be doing

• Take seven products with seven audiences to market without pretending they are one brand
• Own the store listings — three languages, two stores, and markets that rank on different things
• Work on the product pages themselves, which are where almost every decision about a product gets explained

What we would expect

• You have launched an app in at least one of these markets and know what actually moves installs there
• Arabic or Hindi as a working language, for the UAE and India postings
• You write plainly. Every product here is full of words, and most of them are the marketing.

A bad fit if: Growth hacking, dark patterns, or paid acquisition as the only lever. There is no ad tech anywhere in this portfolio and no data to broker, which is a constraint on this role as much as on the engineering.