WhatNext. went live on the App Store yesterday, August 14, 2026. It's a movie recommendation app, and the thing I actually want to talk about isn't the app itself so much as the decision underneath it: the recommendation engine runs entirely on your device, and it always has.
That wasn't the easy path. It would have been simpler to build a server, log every rating, and let a shared model improve across all users at once. That's how most recommendation products work, and for good reason: centralizing the data makes the model better, faster, with less engineering effort. WhatNext. doesn't do that. Your ratings stay on your phone by default. There's no account, because there's nothing to attach an account to. If you delete the app, you delete the taste profile with it. There's no copy sitting on a server somewhere.
What's Actually Inside
Setup takes a few minutes. You rate a handful of films you've already seen (enjoyed it, it was okay, not for me), or, if your watch history is thin, you can bootstrap from genres and moods instead.
From there the app computes a taste vector, a few hundred numbers that describe what you tend to gravitate toward. Every film in the catalogue has its own precomputed vector, so ranking your daily feed is just similarity math, run locally, in well under a second.
The feed itself refreshes daily, a curated set of picks instead of an infinite scroll. There's full-text search across title, director, and theme, plus a separate mood search for when you know the feeling you want ("rainy Sunday," "mind-bending," "road trip") but not a title. Save anything to a watchlist, mark it seen, rate it, and the model adjusts. Export your entire library as JSON, CSV, or Markdown whenever you want, or wipe it in one tap.
Why This Mattered Enough to Build It the Hard Way
Most film apps are, underneath the recommendations, advertising surfaces. They want engagement and watch-time and upsells, and none of that requires knowing anything about you personally; it just happens to be profitable to collect it anyway. WhatNext. doesn't have that business model, so it doesn't need that data. Building the recommendation engine to run on-device wasn't a privacy feature bolted on afterward. It's the actual architecture, and it shaped nearly every decision that came after it, including some that were genuinely more annoying to build this way (cold-start behavior for new users with thin watch histories being the obvious one: there's no shared model to lean on when a device is starting from zero).
What's Next
WhatNext. is iOS only for now, with Android in progress; the same on-device design carries over when it ships. This blog is where I'll be writing about what happens next: the mechanics of the recommendation engine itself, decisions made along the way, and, since this is still fundamentally a movie app, recommendations, the same way the app itself makes them, just with room to actually explain why.
If you want to try it, it's live now on the App Store.