Open to senior roles · NZ work rights, no sponsorship needed
I’m Rocky Shi, a senior engineer in Auckland. Go is my main language, plus TypeScript, Rust, Swift and Dart. I own work end to end, from design and tests to release and production support. I lead coding agents the way a tech lead runs a team: clear context, small checkable changes, and no merge without proof.
Native apps for macOS, Windows, Linux and the web, built outside my day job. Each one is released, versioned and maintained, and each was built by directing AI coding agents.
Real-time speech-to-text for the Mac: dictate, type by voice into any app, or caption a call live.
Streams partial results from on-device models (Parakeet and WhisperKit on Core ML) or five cloud providers, with speaker labels, custom vocabulary and LLM clean-up. My most-used app: 71 releases and 580 downloads since February.
12:01Most bugs hide in the assumptions you never wrote down.大多数错误藏在你从未写下的假设里。
12:05So write the assumption first, then the test.所以先写下假设，再写测试。
12:09That one habit saved our release.这个习惯挽救了我们的发布。
Podcasts and recordings with synced transcripts, translations and listening practice.
Transcribes locally on Apple silicon or through cloud providers, translates sentence by sentence, and writes AI briefs (chapters, decisions, action items) whose timestamps link back to the recording.
A floating toolbar for selected text, anywhere on the Mac, with AI built in.
Translate, chat, grammar-check or run custom LLM, JavaScript and pipeline actions on any selection. Chat streams answers and can read the source page or file through tool calls, after you approve them. 30 releases since February.
A native debugging proxy for HTTP, HTTPS and WebSocket traffic.
Per-host HTTPS decryption, HTTP/2, breakpoints, map local/remote and JS scripting. It also runs an MCP server, so AI assistants can inspect and replay captured traffic.
第一章 光与影
Chapter One · Light and Shadow
清晨的光从窗外照进来，落在旧书的纸页上。
Morning light fell across the pages of an old book,
and every word waited to be read again.
Turns scanned PDFs into searchable or cleanly retyped documents, fully offline.
Runs PaddleOCR PP-OCRv6 models through ONNX Runtime on the Mac GPU via Core ML, from Go without cgo. It reads pages in layout order, matches type weight and style, and traces line art into vectors.
Mirrors iPhone, iPad and Android screens over USB, and drives Android from the Mac.
Captures iOS through CoreMediaIO and decodes Android H.264 with VideoToolbox at high refresh rates. Small dex helpers inject input and stream audio, with no APK to install.
Quarterly Review
Revenue grew 18% on the year, driven by
new customers in the draftPacific region.
Account 0042 1187 3320 closed.
A fast PDF reader and editor built for the Mac.
A PDFium engine over FFI with native AppKit surfaces: in-place text and image editing, redaction, forms and signatures, with atomic saves that keep CJK and CMYK content intact. A CLI lets AI agents inspect, search and edit the open workspace.
A menu-bar screenshot tool with scrolling capture, annotation and OCR.
A Flutter UI over native capture code in Swift on macOS and C++ on Windows. It auto-scrolls and stitches long pages, snaps the selection to page elements, and annotates, redacts and runs OCR before you copy.
The Yaak API client rebuilt in pure Go, to stress-test a native UI toolkit.
HTTP, GraphQL, gRPC and WebSocket clients, Git sync and a code editor, all drawn natively and tested headlessly.
Binds a domain to any service on a VPS from the browser, with HTTPS.
One Rust binary: reverse proxy, Cloudflare DNS over OAuth, Let’s Encrypt certificates and an admin UI.
A digital logic workbench, from NAND gates to a programmable 8-bit CPU.
A 100-mission course with automated checks. Built in TypeScript and React, then reimplemented in Rust and WebAssembly.
Agents are fast but unreliable. Most of my work is the system around them: they get the right context, work in small steps, and nothing ships without verification.
Turn a fuzzy ask into small milestones, each small enough to review and test on its own.
Agents get the architecture, patterns, commands and what “done” means up front, in an AGENTS.md or the brief, so they start from facts instead of guesses.
An agent’s change is only accepted with proof: tests shown to fail without the fix, headless UI tests and hands-on runs of the real app. When it cuts corners, I send it back.
Signed, notarized releases with changelogs, followed by production support. The same bar I hold at work.
Beyond using agents to write code, I build AI into the products themselves, both at work and in my own apps.
Nine years of backend work in trading, banking data, news and lending. I care most about correctness: exact results, transactional workflows, and tests from day one.
A land-title search service over a PostGIS replica of ~2.4M NZ titles. Weekly syncs publish in one transaction, so the API never serves stale or partial data.
Go · gRPC · PostGISEvent-driven ingestion across 13 blockchains, OpenSearch ranking for token discovery, and OpenTelemetry that caught regressions before customers did.
Go · Kafka · PostgreSQL · OpenSearchLed the split of a stateful single-node search service into a distributed one. Designed an AI content-mining platform that won a banking client soon after launch.
Go · gRPC · RedisKept flash-news search under 100 ms at production load. Shipped the community module that drove significant user growth.
Go · Elasticsearch · MySQLRust contributions to the GPUI editor, from window management to Vim motions.
University of Auckland · 2025 – 2026
Bootstrapped the architecture of a team-built recipe app (ASP.NET Core backend, React Native client, AWS), led its design discussions and prototyped its AI recommendations.
B.Eng · Nanjing Institute of Technology
Based in Auckland, New Zealand. I hold NZ work rights and need no sponsorship.
