On-Device AI

AI That Never Phones Home

When most apps say "AI-powered," they mean your data travels to a server farm, gets processed, and comes back. Every prompt. Every piece of content. Logged somewhere you can't see.

Postiller's on-device AI works differently. The AI models run directly on your iPhone's Neural Engine. Your content never leaves your device. There's no server to send data to—the intelligence is local.

This isn't a stripped-down compromise. Modern iPhones have dedicated AI hardware that rivals cloud servers for many tasks. We use it.

What Runs On-Device

Not every task needs a frontier model. Postiller uses on-device AI for:

Summarization

When you save a bookmark, on-device AI generates a concise summary. You get the key points without reading the full article—and without that article being sent anywhere.

Semantic Embeddings

Every piece of content gets converted into an embedding—a mathematical representation of its meaning. This powers semantic search, content discovery, and context matching. All computed locally using Apple's NLEmbedding framework.

Content Tagging

On-device AI extracts topic tags from your content. These tags help cluster related content and surface unexpected connections between your bookmarks and ideas.

Title Generation

When you save a bookmark without a clear title, on-device AI suggests one based on the content. Quick, free, and private.

The Technology Behind It

Apple Neural Engine

Every iPhone since the A11 chip (iPhone 8 and later) includes a Neural Engine—specialized hardware designed for machine learning tasks. The latest chips can perform 15+ trillion operations per second.

This isn't marketing fluff. The Neural Engine runs the same types of computations that power cloud AI, just optimized for mobile hardware and power efficiency.

Apple Foundation Models

iOS includes built-in models for natural language understanding. These aren't toys—they're the same models powering features across Apple's ecosystem: search in Notes, email categorization in Mail, photo understanding in Photos.

Postiller leverages these models through Apple's frameworks. No downloads, no setup, no API keys. They're already on your device.

NLEmbedding

Apple's Natural Language framework includes NLEmbedding, which converts text into semantic vectors. These 512-dimensional embeddings capture the meaning of text in a form that enables similarity search.

When you search for "leadership advice" and find an article about "management strategies," that's NLEmbedding understanding that these concepts are related—without ever sending your query to a server.

On-Device vs. Cloud: The Trade-offs

Let's be honest about what on-device AI can and can't do.

Where On-Device Excels

Speed for common tasks — No network round-trip means instant results. Summarization and embedding generation feel immediate.

Privacy guarantee — There's no way for your content to leak because it never leaves your device. This isn't a policy we promise to follow—it's a technical architecture that makes leaks impossible.

Offline capability — Airplane mode? No WiFi? On-device AI keeps working. Your content workflow doesn't stop when connectivity does.

Zero marginal cost — Use it as much as you want. Generate a thousand embeddings. Summarize every article. There's no meter running.

Where Cloud Models Win

Raw capability — GPT-4 and Claude have more parameters, more training data, and more reasoning ability. For complex, nuanced content generation, cloud models produce better results.

Long-form generation — On-device models are optimized for efficiency, not extended generation. A 2,000-word blog post needs a cloud model.

Latest knowledge — Cloud models are updated regularly. On-device models update with iOS releases.

The Hybrid Approach

Postiller doesn't force you to choose. We use on-device AI where it excels and cloud models (via your own API keys) where they're needed.

Automatic routing:

You can override these defaults. Want to use Claude for summarization? Go ahead. Prefer to keep everything on-device even if quality suffers? That's your choice.

The point is flexibility. Use free, private, on-device AI for the bulk of operations. Reserve your API budget for the tasks that truly benefit from frontier models.

What "Free" Really Means

On-device AI has no usage fees because the computation happens on hardware you already own. But "free" deserves some context:

No API costs — You're not paying OpenAI, Anthropic, or anyone else for these operations.

No subscription required — On-device features work without any account setup or recurring payments.

Battery usage — AI inference does use power. It's optimized for efficiency, but heavy usage will impact battery life.

Device requirements — On-device AI works best on newer devices. iPhone 12 and later have the most capable Neural Engines.

The economics are simple: you paid for the AI hardware when you bought your iPhone. Postiller just puts it to work.

Privacy Without Compromise

Most "privacy-focused" apps ask you to trust their policies. They promise not to log your data, not to train on your content, not to sell your information.

On-device AI doesn't require trust. The architecture makes privacy violations technically impossible:

• No server connection for on-device tasks
• No data transmission means nothing to intercept
• No logs because there's no server to log to
• No training on your data because Apple's models are pre-trained

This is privacy by design, not privacy by policy. We can't access your on-device processed content because it never reaches us.

Setting Expectations

On-device AI is genuinely useful, but it's not magic. Here's what to expect:

Summarization quality: Good for extracting key points. Occasionally misses nuance or context that a human (or GPT-4) would catch.

Embedding accuracy: Excellent for finding conceptually similar content. Sometimes clusters by writing style rather than topic—we compensate with topic tag extraction.

Speed: Near-instant for most operations. Complex tasks on older devices may take a second or two.

Consistency: Results are deterministic given the same input. The model doesn't have "good days" and "bad days."

For content generation—the task that benefits most from AI quality—we recommend using cloud models via BYOK. On-device handles the supporting infrastructure; cloud handles the creative heavy lifting.

The Future Is Local

On-device AI is getting better fast. Each iOS release brings improved models and new capabilities. The gap between local and cloud narrows every year.

We've built Postiller to take advantage of this trend. As Apple's on-device capabilities improve, more of your workflow can stay completely private—without sacrificing quality.

Today, it's summarization, embeddings, and tagging. Tomorrow, it might be everything.