How DailyVox
actually works.

Every piece of data in DailyVox flows through a pipeline that runs entirely on the device. There are zero network calls for AI processing. Here is the full system architecture.

The key constraint: data never leaves the device for processing. Transcription runs on the Neural Engine. NLP runs locally. The Twin is computed and stored in Core Data. The only optional network path is Apple's encrypted iCloud sync — which the user can disable.

On iPhone, DailyVox uses nine Apple frameworks to build a full AI pipeline without any third-party dependencies or server-side processing. The Android build reaches the same result differently, and §03 sets the two side by side.

DailyVox is live on iPhone. A native Android app is in development — not released, and not a wrapper around the iOS one. It is Kotlin and Jetpack Compose, built against the same Twin engine contracts, because the promise that has to survive the port is the one people actually chose the app for: nothing leaves the phone.

That promise turned out to be the hard part, and not for the reason you would guess. Apple hands you a whole on-device AI stack for free. Android does not — there is no system sentiment API, no system name recogniser, and the strong open NER models inherit training corpora that restrict commercial use. So every gap had to be closed with something measurable rather than something convenient.

Two numbers we had to earn, not assume

Replacing a first-party framework with your own code is only defensible if you measure the replacement against the thing it replaces. Both substitutions were scored on the same real diary text, transcribed from actual recordings.

Sentiment. On 1,459 human-labelled entries, the VADER lexicon scored r = +0.663 with 87.1% sign accuracy, against NLTagger's +0.594 and 79.8%. Both shuffled controls collapsed to chance. The lexicon is not a compromise; on this register it reads the writer more accurately than the OS does.

Names. On 28 transcribed entries with 111 hand-annotated spans, the heuristic found 99.1% of them at 61.6% precision, against NLTagger's 94.6% at 32.5%. Every one of the 43 person names in that corpus is non-Anglo, and the heuristic found 100% of them where Apple's model found 97.7% — because it has no learned prior about what a name looks like, so it has none to be wrong about.

Neither number is a marketing claim. The corpus is one author's diary, 28 entries is small enough that a handful of spans moves a percentage point, and the person who wrote the heuristic also wrote its gold labels. Both are stated here with those limits attached because that is what makes them checkable.

What the port costs, stated plainly

Android's on-device recogniser needs a language pack that this app cannot download — it holds no internet permission of any kind, which is the entire point. On a phone without the pack, recording fails and the app says so, and points at the Android setting that fixes it. It does not quietly transcribe over a network instead.

There is also no free-form conversational Twin on Android. A model small enough to ship to every phone in the target audience would guess, and a Twin that guesses about your own life is worse than one that stays quiet. Ask answers structured questions from real statistics and cites the entries it used — which works on every device rather than only the newest ones.

The TwinEngine is a custom personality-modeling system that builds a multi-dimensional profile of the user from their voice journal entries. It uses no external models or APIs. The entire model is computed from NLTagger output and stored as serialized JSON in Core Data's AIState entity.

The engine consists of four interconnected sub-models.

Where DailyVox has been, what's being built now, and where it's going. Each version adds a layer to the on-device AI stack. Shipped items are live on the App Store; everything below the line is honest about not being finished. The full roadmap has the detail.

DailyVox exists at the intersection of on-device LLMs, personal AI, and mental-health technology. Several recent research papers explore adjacent ideas.

1. [1] 2026
   Memory-Efficient Structured Backpropagation for On-Device LLM Fine-Tuning

   Efficient fine-tuning under mobile memory constraints — the path back to weight-level personalization if on-device adapter training becomes practical.
2. [2] 2025
   MoPHES: On-Device LLMs for Mobile Psychological Health

   Using on-device LLMs for psychological health applications — parallel to the Twin's emotional modelling.
3. [3] 2024
   PocketLLM: On-Device Fine-Tuning for Personalized LLMs

   Personal model adaptation on mobile hardware — foundational for v2.1's personal adapter.
4. [4] 2023
   PLMM: Personal Large Language Models on Mobile Devices

   Early architecture proposal for personal LLMs running on phones — the direction DailyVox pursues.

The DailyVox app is open source under the MIT License — the SwiftUI interface, recording and transcription pipeline, Core Data stack, and all app-side processing are on GitHub. The Digital Twin engine is a separate proprietary Swift Package; the app’s zero-network behavior is verifiable regardless: run it in airplane mode or packet-capture it.

Privacy-critical software should be auditable. If you claim data never leaves the device, people should be able to verify that claim by reading the code.