Two years ago I was writing CRUD endpoints. Now I ship RAG systems, agents, and streaming chat to real users.
Based in Melbourne. Solo-shipped to the App Store. Currently obsessed with closing the gap between LLM demos and shipped products.
Field Notes
May · MMXXVI
Three threads in parallel — vibe-coding with Claude Code, RAG pipelines end-to-end (chunking, hybrid search, re-ranking), and AI agents (tool use, state, when to stop). Less demo-magic, more "why does this actually hold up in production".
Claude Code + Obsidian — the agent-on-your-notes workflow Andrej Karpathy calls a "second brain". Also tinkering with GPT-Image-2 for visual experiments.
"Memory is finite. Externalizing knowledge is how thought scales."
Vibe-coding tiny browser tools with Claude Code — each one a weekend from idea to install. EchoReply (Gmail intent decoder) and JD Analyzer (resume ↔ JD match) just shipped to the Chrome Web Store. Next one already on the bench.
— See the lineup→Always under
construction.
I started in CRUD apps — the kind of work where the hard part is naming things and the easy part is the framework. The pattern got familiar fast. When LLMs started actually working in production, I went chasing what wasn't familiar yet, and I haven't stopped — retrieval, streaming, agents, vibe-coding workflows, the messy second half of an AI feature where the demo ends and the user shows up.
RAG was the first thing I picked to go deep on, because it's the part that decides whether an LLM product feels smart or stupid, and almost no one writes about it honestly. DocuMind is my running notebook on the topic — I rebuild the same pieces (chunking, eval sets, hybrid search) until I can say why each one is the size and shape it is. Agents and vibe-coding are next.
Mostly I care about shipping. A demo I can't put in front of a real user is a research project with better lighting. The work I'm proud of is on the App Store, on Vercel, and now in production as the internal payroll system for Shangxue Tutoring (Mar – Jun 2026) — getting hit by traffic from people who don't know me.
I keep picking up new ways of working — the stack moves, you move with it. Right now I'm vibe-coding AI Chrome extensions with Claude Code. EchoReply is the first one out the door — an AI sidebar for Gmail that decodes what the sender actually wants and drafts replies in three tones, now live on the Chrome Web Store. Proof I'm not just reading about this stuff, I'm shipping with it.
A handful of things I've shipped — each link is a real product or a working repo.
AI-powered SaaS platform — upload PDFs and query them in natural language via RAG. Built with LangChain, OpenAI GPT-4o-mini, and Pinecone vector DB. Streaming SSE responses, source citations, multi-turn memory, and per-user namespace isolation.
Streaming SSE · RAG pipeline · Pinecone vector search · Deployed Vercel + Render
LangGraph-powered learning-path generator. Three compiled StateGraphs (planner / quiz / retry) design a week-by-week roadmap, fetch real resources via Tavily, generate quizzes, and adapt to your weak spots. Wrong answers get embedded into Pinecone with concept tags, so memory shapes future roadmaps — even across different topics.
LangGraph StateGraphs · Pinecone long-term memory · SSE streaming agent log · Adaptive retry loop · Google OAuth + JWT · 15 vitest specs
Internship project for Shangxue Tutoring — production payroll platform replacing paper-based teacher salary claims. 10+ weeks shipping with one PM and Claude Code as my AI pair. Three-tier RBAC (teacher · admin · superadmin), invitation-based onboarding, live AUD→CNY rate locked per claim, bilingual EN/Chinese UI. Layered Express backend with 39 unit tests.
Internship (Mar–Jun 2026) · 10+ weeks shipping · 3-tier RBAC · Bilingual · 39 backend tests
AI email coach that lives in Gmail. Open any thread and the sidebar reads the sender's real intent, emotional temperature, and reply urgency, then drafts replies in three voices — Professional / Warm / Confident — plus a Tone Check that audits your own drafts before you hit send. Powered by your own Claude or OpenAI key — emails never touch a third-party server.
Chrome Web Store · Manifest V3 · Vanilla JS, zero build, zero deps · 24h local cache · Multi-provider (Claude · OpenAI) · Zero backend
Privacy-first Chrome extension that scores your resume against any job description in seconds. Open a posting on LinkedIn, Seek, or Indeed, click Auto-fill, and get a 4-dimension match score, skill-gap analysis, streaming cover letter, line-by-line resume rewrites, and 10 predicted interview questions. Powered by your own Claude or OpenAI API key — every byte stays in your browser.
Live on Chrome Web Store (May 2026) · 8 job sites · ~$0.01–0.03 per analysis · 37 unit tests · Zero backend, zero tracking
Bilingual six-chapter Chinese-language guide for Melbourne motorcycle newcomers — VIC licensing, bike picking (11 curated + 43-bike catalog with side-by-side compare), gear, 12-month skill progression, Black Spur route notes, and insurance comparison. React 18 + Vite SPA pre-rendered to static HTML with Puppeteer so each chapter is SEO-indexable, content driven by a single editable data file.
6 chapters · 43-bike catalog · Puppeteer SPA pre-rendering · React 18 + Vite · Deployed Vercel
Full-featured cloud storage and file-sharing platform. Passwordless email OTP auth, real-time data sync, and modular architecture built for scale. Next.js 15 App Router with TypeScript and AppWrite backend.
Passwordless OTP auth · Real-time sync · Next.js 15 · Deployed Vercel
Notes on building, shipping, and what I'm learning along the way.
Vibe coding isn't 'AI writes my app while I nap.' It's a tighter loop — brief well, read the diff, verify the behaviour — and Claude Code is the closest thing to a competent junior I've worked with.
Why I shipped a privacy-first Chrome extension that reads sender intent, drafts replies in three voices, and audits your own tone before you hit send — and what 'no backend' actually means in practice.
Why I shipped a Chrome extension that scores resumes against job descriptions client-side, on the user's own AI key — and what the architecture had to look like to make 'no backend' actually true.
10 weeks shipping an internal payroll platform for a tutoring company — paired with a PM and Claude Code as my AI pair. The decisions worth defending and what changes about the work when AI sits at the keyboard with you.
Open to full-stack and AI engineering roles. If you're building something where shipping matters, drop me a line.
taoaaron5@gmail.com
Replies within 24h · Australian PR
