Nicholas Fong

Built a native macOS app in Swift that remembers which app and window you were in, and when.

• Reads the app in front and its window title through macOS Accessibility. No screenshots, screen recording or audio.
• Groups your day into moments on a timeline you can search.
• AI apps like Claude and Cursor connect in one click, then ask it questions through a read-only MCP server with no network port.

Built a shopping assistant where several AI agents researched products in parallel and compared them.

• A discovery agent pulled candidates from expert reviews, parallel agents researched each one, and a synthesis agent ranked them, streaming every step live.

150+ MAU at peak.

Built an AI agent for Soma Capital’s firm-wide chat that posted a morning recap to all ~31 people and answered tagged questions from the firm’s email, calendars, CRM and code.

Within a month it added email triage for the Founding GP, warm-intro research, people search and meeting prep.

Used daily by the entire firm (~31 people).

Built a personal AI assistant that runs on a Mac mini and answers from one local index of texts, email, photos, files and calendar.

• Every claim links to the exact message it came from, and replies go only into existing threads, after two presses.
• Remote browsers: each chat drives its own tabs in a headless Chrome on the Mac mini, streamed live to laptop and phone; from the laptop you can take over at any point.
• Stateful agents to reduce the effects of context compaction: one live session per chat, resumed after restarts, and only what changed sent each turn.

Fine-tuned Qwen QwQ-32B, an open reasoning model, using 10k synthetically generated reasoning traces to make it think in a hierarchical pattern.

Kept its accuracy with ~80% fewer tokens on similar tasks; on tasks with a different composition structure, accuracy dropped significantly.