Software you shouldn't have to trust

A small confession to start. Every time I install an app and tap “Agree,” some part of me thinks about everything I’ve just agreed to, and how little of it I actually read. We all do it. We hand over our photos, our messages, our health history, our private writing, to companies we’ve never met, and we trust them not to look at it, lose it, sell it, or be made to hand it over.

For the last couple of years I’ve been quietly building software on the opposite premise. Not “trust me with your data.” Instead: what if you never had to?

The bargain we stopped noticing

The arrangement was always a little uncomfortable, but consumer AI made it steeper. The genuinely useful new tools are overwhelmingly cloud tools. Your camera roll, your inbox, your medical records, your half-finished thoughts, streamed to a data centre to be processed and, very often, kept to train the next model. The convenience is real. So is the quiet cost: a whole generation of deeply personal software that only works by sending your most private material somewhere you can’t see.

I didn’t want to build more of that. So the studio I run, Purvakalpa, starts from a refusal.

The thesis, in four lines

• Native Apple platforms.
• The AI runs on your own hardware.
• Nothing leaves your iCloud.
• The cluster of devices you already own is the backend.

Each line is a deliberate choice.

Native means no shortcuts. Every app is built directly on Apple’s frameworks, in Swift, because that’s the only way to reach the on-device machine learning and the privacy hardware that make the rest possible.

On-device means the intelligence runs where you already are. Small AI models run locally on the phone; Apple’s own on-device vision and language models do the heavy lifting. When a cloud model genuinely helps, it’s strictly opt-in, your own account, your own key, and you decide, feature by feature, whether anything leaves the device at all.

No servers means there is no Purvakalpa backend. No central database of users to breach, no logs to subpoena, no infrastructure bill that has to be paid by quietly monetising your attention. Your data lives in your own iCloud, or stays on your device. I never see it. That isn’t a promise of restraint; it’s the simple absence of a mechanism.

And the last line is the part almost nobody builds. Most of us own three or four Apple devices sitting idle next to each other. Treated as what they are, a small, private compute cluster, your phone can hand a heavy job to your Mac and get the answer back over your home Wi-Fi in under a second. No cloud required, because the supercomputer was already in the room.

Six apps, one stubborn idea

A thesis is only worth what it can build, so I’ve deliberately built across categories that share nothing except the conviction underneath:

• PawHub, household pet care, live on the App Store. Coordinates a whole family through Apple’s sharing, with the sensitive records kept in a store wired so it can’t sync anywhere.
• Knight Fever, a chess game, live on the App Store. No accounts, no servers; peer-to-peer play with player identities cryptographically sealed before they ever touch the network, and a full play-by-voice accessibility pass.
• Volley, a cricket coaching tool, in development. Watches practice footage on-device and learns each coach’s eye over time, with no clip ever leaving the local network.
• Unknot, a private mail client for the Mac, in alpha. Reads and summarises your inbox on-device, with the mail itself held in a store that’s architecturally incapable of leaving the machine.
• Kokoro, a wellbeing companion for iPhone and Apple Watch. Before its AI is allowed to answer a question about your health data, a rule engine answers it from your real numbers, so it can’t invent them. Your health data never leaves the device.
• Storyline, a writing app, and the engine beneath everything. A runtime that discovers your other devices on your network, confirms they’re yours, and shares the AI work between them.

Look closely and they’re really one platform wearing six faces: the same privacy split, the same on-device AI, the same encryption, reused again and again. More than seven hundred files of it now, and the most-reused parts are precisely the ones that protect the person using them.

The part that actually earns trust

Conviction is cheap, so here’s what I try to hold myself to.

Choose the harder mechanism over the easier promise. “Local-only,” in these apps, isn’t a setting that asks you to believe me; it’s a data store configured so the system itself can’t upload it. The trust anchor is an impossibility, not an intention, and an impossibility can’t change its mind after an acquisition.

Correct yourself in public. One app’s promise used to read “nothing leaves this Mac,” until I realised a small key and a summary do leave, into your own iCloud, never a vendor, and rewrote it to the exact truth: “nothing leaves your iCloud.” A cleaner tagline wasn’t worth being slightly wrong.

And ship. Two of these are already on the App Store, through Apple’s review and its privacy labels. This isn’t a manifesto waiting for funding; it’s running on real devices today.

Why now

For most of the last decade, this wasn’t buildable. The models were too big and the phones too slow, so the cloud was less a choice than a necessity. That has quietly reversed. Apple silicon put a real neural engine in every device, small models got genuinely good, and the on-device frameworks arrived. The cloud-only assumption baked into most AI products is now a choice, not a constraint, which means people can finally be offered the other option.

So, honestly, will this help anyone?

That’s the part I don’t know yet. I’m fairly sure I’m not the only person who feels a small unease about how much of my life lives on other people’s servers. But “fairly sure” isn’t the same as right.

So I’ll just ask. If you’ve ever hesitated before handing an app your photos, your messages, or your health data, does an approach like this resonate? Would you choose the version that keeps your life on your own hardware, even if it asked a little more of you? Or am I solving a problem that only I have?

I’d genuinely value your honest answer, even, especially, if it’s “no.”

Eswar Venigalla, Purvakalpa (Bengaluru)