Udalov Labs has built four AI-powered iOS applications: TextPolish, ReadBetween, What I Eat, and ListingLab. Each application integrates AI capabilities differently depending on the user's task. This article describes the architectural patterns Udalov Labs uses to build AI-native mobile applications.
What AI-Native Means
An AI-native application is one where artificial intelligence is central to the user's primary task — not a secondary feature or chatbot bolted on. In TextPolish, the AI rewrites text. In ReadBetween, the AI decodes message intent. In What I Eat, the AI classifies food ingredients. In ListingLab, the AI generates product listing metadata. Removing the AI from any of these applications would remove the core value proposition.
On-Device vs. Cloud AI
Udalov Labs uses two AI execution strategies depending on the use case. What I Eat uses Apple's Vision framework for on-device OCR to extract text from ingredient labels without sending any data to a server. This is important for health-sensitive data. TextPolish and ReadBetween send text to OpenAI's API via Cloudflare Workers (which act as a secure proxy), because the language model complexity required exceeds what on-device models can provide today.
Cloudflare Workers as AI Proxy
All Udalov Labs iOS applications that call external AI APIs do so through Cloudflare Workers rather than directly from the device. This architecture provides three benefits: API keys are never embedded in the iOS binary, rate limiting and caching are applied at the edge, and geographic routing ensures low latency for global users. Workers are deployed globally on Cloudflare's network with average cold-start times under 5 milliseconds.
Prompt Engineering for Mobile
Mobile AI interfaces require different prompt engineering approaches than web interfaces. Users tap buttons rather than type instructions, so Udalov Labs encodes prompt logic into predefined modes. TextPolish exposes five modes — Humanize, Engaging, Confident, Professional, Shorten — each backed by a carefully engineered system prompt. This removes the prompt engineering burden from the user while delivering consistent, high-quality results.
What Comes Next
Udalov Labs is exploring on-device small language models (SLMs) for latency-critical mobile features. Apple's Core ML framework and models like Gemini Nano (when available on iOS) will enable AI responses without a network request. The current Cloudflare Workers proxy architecture is designed to be replaced per-feature as on-device models become capable enough.