Why I believe in local AI on phones
I keep shipping consumer apps where the AI runs on the phone itself. Ragondin Assistant is an assistant with no cloud behind it. Recall records and transcribes in-person conversations without the audio ever leaving the device. This is not a research position or an aesthetic preference. It is a product bet, and I keep doubling down on it for three reasons.
A promise is not an architecture
Most privacy in software is policy. “We don’t sell your data” is a sentence in a document, written by people who can be acquired, breached, subpoenaed, or replaced by a growth team having a bad quarter. Every policy is one decision away from being a different policy. Users know this, even when they cannot articulate it; it is why they hesitate before typing the real question.
When there is no server, the entire category of question disappears. Data cannot leak from where it never went. “No data collection” stops being a promise you have to trust and becomes a property of the system, checkable by anyone who watches the network traffic.
This matters more for assistants than for almost any other software, because of what people actually say to them. The conversations are about health, money, marriages, conflicts at work. A recording of your meetings is more sensitive still. This is the most personal data a person produces, and the honest place for it is the device in their pocket.
The phones got good enough
This bet was not available five years ago. Today a mid-range iPhone runs a 2-billion-parameter model at conversational speed, a high-end one runs far larger, and Apple ships its own model inside the OS. For the daily duties of an assistant, answering, summarizing, drafting, transcribing, on-device models crossed the usable line a while ago, and every hardware generation moves it.
The economics nobody talks about
On-device inference costs the developer nothing per user. There is no API bill that scales with success, no server fleet to keep warm. That changes what is possible to ship: apps that are genuinely free or paid once, instead of subscriptions that exist to feed an inference meter. It changes who can ship: a solo builder can serve a million users with the same infrastructure as ten, which is none. And it changes where things work: in airplane mode, in the subway, behind any firewall, with no round trip to a data center before the first word appears.
What local AI is bad at
Being clear about the boundary is part of the bet. Small models do not do frontier reasoning. They will not analyze a hundred-page contract, write serious code, or know what happened yesterday. When I need those things, I use a frontier model in the cloud like everyone else. Local AI does not replace that layer and should not try.
The territory it wins is different: the private, personal, always-available layer. The question you would not type into a service. The meeting you want transcribed but not uploaded. The assistant that is still there with no signal and no subscription. Those jobs are small enough for a model that fits in your pocket, and private enough that they should never have to leave it.
The bet, stated plainly
Privacy that is architecture instead of policy, on hardware that is finally good enough, at a marginal cost of zero. Each of those alone is an interesting feature. Together they are a category, and it is early enough that the defaults are still being decided. That is why I keep building here.
Making all of this invisible to the user is real engineering: choosing models per device, moving gigabytes of weights onto phones, surviving iOS memory limits. I wrote up how that actually works at the studio blog: Shipping an on-device LLM in a consumer iPhone app. This post is the why; that one is the how.