On-device and offline AI for private mobile apps

The problem: useful AI often needs private data

The most useful AI features work on sensitive material: personal chats, customer records, medical notes, contracts. Sending that to a cloud model raises privacy concerns, legal questions and trust issues. And many Indian field-work settings (villages, warehouses, basements, remote sites) simply have no reliable connectivity.

On-device AI avoids both problems. The model runs on the user's phone. Data stays there, and the feature works in airplane mode.

Our own shipped example: Kepton

Kepton is our on-device Android app, live on Google Play. It drafts replies in the user's own voice and shows conversation pattern insights. There is no account, and the only network step is a one-time model download. It shows that this approach works on real phones, not only in demos. Read its privacy policy and terms to see how a privacy-first app describes itself.

How an on-device AI app is built

1. Choose the task. Drafting, summarising, classifying, extracting or answering questions over local documents.
2. Choose the model. A small language model sized for phones, often quantised (compressed) to fit memory. For search, a compact embedding model turns text into vectors.
3. Store locally. Documents, chats and vectors sit in an encrypted database on the device.
4. Retrieve and generate. The app finds relevant local passages and gives them to the model, a local version of RAG.
5. Run in the background safely. Inference runs in a managed service with limits on memory, heat and battery.
6. Deliver the model. The model file downloads once on Wi-Fi, then the app stays offline. Updates ship as new model versions.

Inputs, what the AI does, outputs

Input	What the AI does	Output
Chat history on the phone	Learns tone and drafts replies locally	Private reply suggestions
Field forms, notes, manuals saved offline	Answers questions and fills fields without a network	Field-staff assistant that works offline
Personal or business documents	Embeds them and searches by meaning	Private document Q&A with cited passages
Voice or text notes	Summarises and classifies on the device	Structured notes ready to sync later

Trade-offs to know about

• Model size. Phone-sized models are smaller than cloud models, so they are weaker at complex reasoning. They suit focused tasks very well.
• Storage and download. Models take hundreds of megabytes to a few gigabytes. A one-time download over Wi-Fi is normal.
• Battery and heat. Long generation uses power. We tune context length and response length to keep it reasonable.
• Device range. Older or low-RAM phones may be too slow. We define a minimum device spec and test on a real range of Android phones.
• Updates. Improving the model means shipping a new version to users, not changing a server.

When these limits matter more than privacy, a hybrid design can use the cloud for hard tasks with the user's consent. We help you decide where the line should sit.

Integrations

Offline apps can sync with your systems when a connection returns: a CRM, an ERP or a custom backend. We build the mobile app and the sync layer as custom software, and the AI features with generative AI expertise.

Guardrails and human review

• The user sees and approves any draft before it is sent.
• Local data is encrypted and the app asks only for permissions it needs.
• The app states plainly what stays on the device and what, if anything, leaves it.
• Output quality is tested on real phones and real content before release.

Privacy law still applies to on-device apps in some ways, for instance on notices and consent. Confirm with your adviser.

Data you need to start

• The one task you want the app to do, with 50 or more realistic examples of inputs and good outputs.
• Your target device range: minimum Android version, RAM and typical handsets of your users.
• Any documents or forms that must be available offline.
• Your privacy requirements: what may never leave the phone.

A hypothetical walkthrough

Picture a field technician in a basement plant room with no signal. She opens the app and asks, in plain words, how to reset a particular pump controller. The app embeds her question, finds the matching passage in the service manual stored on the phone, and the local model writes a short answer that cites the page. She logs the visit by voice; the app turns it into a structured note and holds it until the phone is back online, then syncs. No manual text left the device for the answer. The scenario is invented.

What changes for your team day to day

Staff keep working where signal is poor, and nobody has to decide whether it is safe to paste sensitive text into a cloud tool. For you, support shifts from running servers to managing app and model releases, and watching device compatibility reports.

Rollout

We begin by choosing a single task and testing two or three candidate models on real devices from your target range. A prototype app follows, so you see speed, battery use and answer quality before committing. Then comes the full build, a closed beta and release on Google Play. After launch, we handle model updates and device-compatibility fixes under an annual maintenance contract.

Who it fits

Businesses with sensitive data, field teams without reliable connectivity, and founders building privacy-first consumer apps. See Kepton for a live example, or talk to us about yours.