AI Mobile App Development · On-device & private
We build AI into iOS and Android apps the hard way — on the device, where it's private, instant, and works offline. Core ML and Apple Foundation Models on iOS, Gemini Nano and ML Kit GenAI on Android, and on-device LLMs where you need them, with a cloud fallback only when the model can't fit. First working build in 4–6 weeks, from $8K.
Who we build for
The build decision
There are three ways to put AI in a mobile app. A cloud AI API (OpenAI, Gemini, Claude) is fastest to ship but sends user data off the device and stops working offline. A platform on-device SDK (Apple's Core ML and Foundation Models, Google's ML Kit GenAI on Gemini Nano) runs locally and privately, but only on supported hardware and within a fixed feature set. A custom on-device build tunes the model to your data, runs on the devices you actually support, and is the only option when accuracy on your domain, privacy, or offline use is the product. Here's the honest trade-off.
The pipeline
On-device AI replaces the cloud round-trip with a model that lives inside the app binary and runs on the phone's NPU. Here's the path from picking a model to shipping it in the App Store and Play Store — and where the engineering actually is.
Figure 1: On-device AI mobile pipeline — model selection to shipped app, with where the work lives.
We pick the model that fits the job and the device: a platform model (Apple Foundation Models, Gemini Nano) when it covers the task, or an open model (Whisper, a small LLM, a vision model) when you need control.
accuracy × size × NPUWe convert to the on-device runtime — Core ML on iOS, LiteRT (TensorFlow Lite) or ONNX Runtime on Android — and quantize (8-bit/4-bit) so the model fits in RAM and runs on the Neural Engine or NNAPI without draining the battery.
the hard partThe model runs locally through Core ML / NNAPI / GPU delegates. No data leaves the phone, no round-trip, works in airplane mode. WhisperKit for speech, MediaPipe for vision, llama.cpp or MLC for LLMs where the platform model won't do.
0 network hopsWe wire inference into the app — Swift/SwiftUI on iOS, Kotlin/Compose on Android — with the model loaded lazily, memory managed, and the UI responsive while the NPU works.
Swift / KotlinA graceful cloud fallback (Apple Private Cloud Compute, or your endpoint) for devices that can't run the model, and an over-the-air model-update path so you ship a better model without an app-store release.
cloud only when neededA tuned on-device build answers in tens of milliseconds, fully offline, with nothing leaving the phone — the combination a cloud API can't give you. The hard part isn't the demo; it's making the model fit, stay fast, and not melt the battery across the device range you support. For the framework-level detail, see how Core ML powers on-device AI on iOS.
Why now
For years, “AI in a mobile app” meant a call to someone else's cloud. 2026 is the year that flipped. Three things made on-device the default.
Apple's Foundation Models framework (introduced at WWDC 2025, expanded at WWDC 2026 with image input, on-device Vision, and a free Private Cloud Compute tier) puts a tuned LLM in every recent iPhone. Google's ML Kit GenAI APIs run Gemini Nano on-device through AICore, now on the Pixel 10 line with Gemini Nano 4 in developer preview.
Modern phones ship Neural Engines and NPUs fast enough to run real models locally; the flagship RAM floor (12GB+) makes on-device LLMs practical, not a demo.
Health, finance, and enterprise buyers increasingly require that data never leaves the device. On-device AI is how you ship those features at all — and Apple is open-sourcing the Foundation Models framework in summer 2026, deepening the toolchain.
Being early is the advantage. The teams that ship correct on-device AI in 2026 own the use cases cloud APIs can't touch — private health, offline field tools, instant camera and voice features — before the field crowds in. We've spent twenty years on the hard half of mobile: the real-time pipelines, the memory budgets, the battery, the device fragmentation. The AI model is new; the engineering that makes it run on a phone is what we've always done.
What we build
Object, garment, document, and scene recognition with custom models (YOLOv8m, CLIP) on Core ML and TensorFlow Lite, at frame rate, no server. We built the AI Wardrobe App: it recognizes garment type, fabric, color, and cut on-device and suggests outfits. Same engineering as our on-device computer vision work.
Analysis the cloud can't touch. We built emoproof — on-device Core ML facial-emotion and voice-sentiment analysis that journals mood over time, with nothing leaving the device.
On-device video effects at frame rate on the phone's GPU and Neural Engine. We built SuperPowerFX (500K+ downloads, 4.6★) and Anime Power FX, featured by Apple in “New apps we love.” Need AI video that scales to large live audiences? See scalable AI video.
On-device speech-to-text with WhisperKit, refined into publish-ready output. We built vBoard, an Android AI voice keyboard. Need it as a full on-device speech-to-text system? That's a sibling build.
Frame-by-frame video analysis on the device, even with the screen locked. We built Golf Tracking App — multi-angle swing and ball-flight tracking on iOS.
Summarize, rewrite, classify, and answer over the user's own data — on-device with Gemini Nano or Apple Foundation Models, or a quantized small LLM (llama.cpp, MLC) when you need full control. For conversational AI video agents, see our pillar work.
When custom wins
A cloud AI API or a platform SDK is the right call when its feature set fits and you're happy renting the intelligence. Custom wins when accuracy on your data is the product, when privacy or offline use is non-negotiable, or when you need the feature to run on the devices your users actually carry — not just this year's flagships. It wins at any audience size — a thousand users or ten million.
Figure 2: Build vs Buy — privacy & on-device requirement × control and future-proofing. Custom wins the top-right at any audience size.
How we work
An idea and a target device, no app yet. We pick the on-device model and runtime, build the iOS/Android app around it, and ship a working AI feature.
Add AIYou have a live app and want a private, on-device AI feature in it. We integrate the model, manage the memory and battery budget, and ship without a rewrite.
TakeoversYou inherited a half-built AI app or a cloud feature that's too slow, too costly, or too leaky. We move the critical path on-device, stabilize it, and extend it.
Pricing
Ballpark starting figures, not quotes. Final scope depends on platforms (iOS, Android, or both), the model, device coverage, and accuracy targets — run the calculator for an instant estimate.
Two ways to work with us. Most clients start on time and materials. If you need one number signed off, we fix it once the scope is documented.
How most of our clients work
When you need one number signed off
Free for qualified projects
Before any contract, we'll give you something useful. Pick the one that fits where you are.
Competitor analysis, core feature definition, monetization modeling, and a full launch blueprint — delivered within a week. Written by engineers who'll build what they plan.
An independent review of your system's technology choices, structural components, and workload fit — with a plain verdict on what's working, what's a liability, and exactly what to change to reach your goal. Delivered within a week.
A full audit of your code with every issue documented, evidenced, and located — exact file, exact line. Plus a system architecture review and a prioritized fix roadmap. Not a consultant's opinion. A case file. Delivered within a week.
A specialist review of your video or streaming product covering latency, media server architecture, WebRTC, playback reliability, real-time chat, and scalability. Every finding is specific, located, and fixable. Delivered within a week.
Why Fora Soft
Twenty years of iOS and Android apps where latency, memory, and battery are the whole game — the exact constraints on-device AI lives inside.
Core ML, Apple Foundation Models, Gemini Nano / ML Kit GenAI, LiteRT (TensorFlow Lite), ONNX Runtime, MediaPipe, WhisperKit, llama.cpp — in real apps. AI Wardrobe App (YOLOv8m + CLIP on-device), emoproof (Core ML), SuperPowerFX (500K+ downloads), AI Keyboard App (Whisper).
We build for the buyers who can't use the cloud — health, finance, enterprise — where on-device is the only way the feature ships at all.
Senior engineers, no offshore handoffs, 250+ products since 2005, and a 100% job-success score on Upwork. We finish and hand over clean.
FAQ
Can AI really run on the phone, or does it need the cloud?
Why build on-device instead of calling a cloud API?
Which on-device frameworks do you use?
Will it work on older phones?
Is on-device AI good for HIPAA / GDPR apps?
iOS, Android, or both?
Can you add an AI feature to our existing app?
How do you update the model after launch?
How much does an AI mobile build cost?
How long does it take?
Keep reading
How Core ML powers on-device AI on iOS
Speech recognition with neural networks, on the phone →Knowledge BaseEngineering AI into app features
The playbook for AI in social & UGC apps →ToolEstimate your build
Instant ballpark on scope and cost →Knowledge BaseWhere AI meets video engineering
The course this sits inside →Tell us the feature, your target devices, and your privacy constraints. We'll come back with an on-device-vs-cloud recommendation and a realistic plan — usually within a day.
Within 24 hours you get a realistic estimate, technical recommendations and next steps. Free, no obligations.
