AI-Native Engineering · 2 yrs
AI-native engineering on real products
About two years of daily multi-agent work on personal products across iOS, Android, Web, and Mac.
- Direct models and tools — then own the merge, security, and ship.
- Real products: Ohga, Hone, DesignOS, Mouse Utility Pro — not tutorials.
- Same bar on every surface: iOS, Android, Web, Mac.
- Looking for full-time team work at any solid level.
- Open to full-time roles where AI, product ownership, and robust systems meet. Open roles.
Insight
Engineering loop mix
Rough share of AI-assisted build time — practiced daily on personal products.
iOS · Android · Web · Mac
Four shipping surfaces, one ownership standard. Toolchains differ; review discipline does not.
iOS
Swift · SwiftUI · XcodeNative product surfaces where agents accelerate UI flows, networking, and test scaffolding — then humans own App Store constraints, privacy manifests, and device feel. Simulator-first, physical device before ship.
Ohga wellness (native tracks) · macOS-adjacent product loops
Xcode + agent pair for SwiftUI; never auto-merge without device pass.
Android
Flutter · Kotlin-aware · GradleCross-platform mobile with Flutter where one agent loop drives iOS and Android parity, plus platform channels when native APIs matter. Play Store, permissions, and performance budgets stay human-owned.
Ohga Live app (Flutter) · Consumer wellness multi-surface
Agents scaffold screens and state; release signing and store review stay manual.
Web
SvelteKit · Next-class · TailwindFull-stack product and marketing systems: dashboards, protocol UIs, studio sites. Fastest agent loop — preview, deploy, instrument — with the highest risk of confident wrong architecture if you skip review.
Portfolio / DesignOS · Independent product surfaces · Protocol and product UIs
MCP + local preview tightens the loop; auth and data models get hand review.
Mac
Swift · AppKit · SwiftUILocal-first utilities and menu-bar products: clipboard intelligence, mouse customization, desktop workflows. Sandbox, keychain, and offline posture matter more than cloud convenience.
Hone Intelligence · Mouse Utility Pro · Buds App
Agents draft AppKit glue; entitlement and privacy claims are non-negotiable.
Where this maps
Useful for full-time seats that want AI leverage with human ownership — any solid level.
Developer / senior
Feature work with agents, MCP, review, then ship.
Lead / staff / principal / architecture
Frame systems, sequence work, own reviews — still hands-on when needed.
Platform, cloud, infra, systems
CI, cloud, ops loops next to product code.
Full-time preferred
Part-time advisory stays limited. Looking for a real team seat.
What I actually run
Not a logo wall — role of each harness in the loop. Languages and cloud live on Stack.
Primary AI-native editor. Multi-file refactors, Composer-class agents, rules, MCP, and parallel agent windows when the change set lives in the repo.
Long-horizon autonomous work: architecture passes, multi-step migrations, deep review. Project instructions, skills, hooks, subagents, and MCP for repo-scoped autonomy.
AGENTS.md-first agent shells for plan–execute loops and PR-shaped work when the repo is instruction-wired.
Spec-driven and agentic CLI workflows for structured planning, implementation passes, and verified delivery loops.
Flow-state completion and enterprise-friendly agent surfaces when the team already lives in GitHub and wants low-friction assist.
Alternate agent IDE for exploration and Cascade-style flows — useful as a second harness opinion on the same change set.
How work gets done
01
Frame before generate
Constraints, success criteria, and non-goals first. Agents fill the frame; they do not invent the product boundary.
02
Small closed loops
Preview, typecheck, and tests in the same session. Prefer verifiable steps over multi-thousand-line dumps.
03
Human owns merge
Diff literacy is the job. Security, auth, data models, and store compliance are never fully delegated.
04
Ship with evidence
Local preview, CI gates, TestFlight / internal tracks, and deploy checks. Close the agent cycle with artifacts, not assumptions.
What I will not outsource to a model
Trust boundaries
Auth, secrets, and data models stay human-owned. Models propose; production policy decides.
Diff literacy over speed theater
Fast generation is worthless if the review is performative. Read the change; reject confident wrongness.
Platform honesty
Store rules, sandbox, and performance budgets are not optional because an agent wrote the PR.
Harness pluralism
No single vendor owns the loop. Cursor, Claude Code, Codex, Copilot, and Kiro-class CLIs each have a role.
Where agent work showed up
Personal products under multi-agent loops
Hone, DesignOS, Ohga, Mouse Utility Pro — shipped with AI-native IDEs and terminal agents under human merge ownership.
Architecture and research acceleration
Architecture drafts, protocol notes, and implementation spikes assisted by harnesses without abandoning judgment.
Team-facing literacy
Fable Skills and writing on RAG, MCP, and instruction files so others adopt AI-assisted engineering deliberately.
Case studies built this way
DesignOS — Local-First AI Design Studio
End-to-end design workspace — onboarding, home, project canvas, design-systems manager, settings/telemetry — without cloud lock-in as the default.
Hone Intelligence — macOS AI Clipboard
Production-grade local-first clipboard product with dual-process AI safety (Swift policy + Rust redaction) and optional org backend.
Ohga — AI Fitness & Wellness Platform
Full multi-surface product: Flutter 3.41 / Dart 3.8, Riverpod, Drift (encrypted SQLite), go_router; Go backend; Next.js admin; Astro marketing site.
Mouse Utility Pro — macOS Mouse Customization
Deterministic input pipeline that keeps capturing if the UI restarts; competitive track vs Logi Options+ documented in-repo.
EngageOS — Gamification Engine
Embeddable engagement infrastructure — points through anti-cheat as a productized engine.
Fable Skills — Agent Skill Library
Machine-readable registry + validated skill packs agents can load for architecture, mobile, design, and platform work.