Mikhail M.

Need an AI assistant inside your iOS or Mac app that does real work, not just chat? I build them: assistants that call your app's functions, answer from your users' data and connect to other services over MCP. Senior iOS developer, 14+ years. I've done this before: a 79-tool AI assistant for macOS, two years and 1,500+ hours with one client. Speech and search run on the device where privacy matters. WHAT I'VE SHIPPED → Aide, an AI assistant for macOS. I planned and built it over two years and 1,500+ hours with one client, from the first version to production. • Tool calling: the model has 79 tools. They cover the user's meetings, calendar, contacts, notes, mail, browser history and local files, plus Google Docs, Sheets, Slides and Drive, Canvas LMS, shell commands and browser control. The app runs each call on the Mac, sends the result back to the model, and repeats until the answer is ready. Risky shell commands ask the user first. • MCP: users connect remote MCP servers with a URL and a token, Zapier included. Their tools show up in the chat with no custom code per integration. • Automations and skills: flows run when a meeting ends or on a schedule and produce summaries, action items, webhooks or MCP calls. Skills are importable SKILL.md instruction packs. The model can draft a new flow or skill, but it is saved only after the user confirms. • RAG: a local vector index in SQLite with an on-device embedding model covers meetings, screen text and past chats. PDFs are searched in OpenAI vector stores with file_search. • Voice and meetings: speech to text on the Mac with WhisperKit, voice activity detection and echo cancellation. Live summaries appear during a call. A speech-to-speech assistant on the OpenAI Realtime API uses the same tools. • OpenAI Responses API with streaming, GPT-5 models, web search and image generation. StoreKit 2 subscriptions and Sparkle updates. → Tempo, an AI health coach for iPhone on the App Store. I built its AI chat on the OpenAI Responses API and 11 function-calling tools that read the user's HealthKit data: sleep, workouts, bloodwork, daily snapshots and the healthspan score. The tools run on the phone and return only their results, so the model never gets the raw health data. I also built on-device posture analysis with MoveNet and Core ML, the HealthKit data layer and the health dashboard. → SHRED, a fitness app with 2.7M users. I built the Apple Watch app in SwiftUI and the StoreKit paywall, added 60 FPS on-device pose detection, and moved Objective-C to Swift without stopping releases. → Scopeburn, my own product: a native macOS app and a React web app on Supabase and PostgreSQL. Live at scopeburn.app. WHAT I CAN DO FOR YOU → Add an AI assistant to your iOS or macOS app. The model calls tools built on your data and backend, and anything that changes data asks the user first. → Connect MCP servers so the assistant can use outside services. → Answer from the user's own data (RAG), on the device or in the cloud. → On-device AI with WhisperKit, Core ML and Vision, for when data shouldn't leave the phone or Mac. → iOS development and takeovers: Swift, SwiftUI, UIKit, Objective-C, App Store releases. HOW I WORK I start with a small first milestone that works end to end, usually two or three tools on your real data. I keep AI behaviour predictable: rules live in code, tool inputs are checked, and answers point back to their source. Tell me which two or three things your assistant should look up or do first. That decides the first tools, and I'll come back with a first milestone and an estimate.