I'm a Melbourne-based full-stack developer focused on AI/LLM applications, AI developer tooling and reliable backend/data workflows.

I build across TypeScript, Node.js, Next.js and Python, with a particular interest in making AI-assisted systems reliable: validating model outputs, separating probabilistic reasoning from deterministic logic, defining clear control and data boundaries, and testing the paths that matter.

• Currently building AI Harness, a model-neutral workflow for AI-assisted software development with GitHub-based NEXT → IMPLEMENT → ACCEPT → SHIP stages, exact-candidate boundaries and explicit human acceptance
• Built DSH Conversation Exporter, a DeepSeek Harness plugin for clean Markdown conversation export and cross-assistant handoff
• Built MoneyGuard, a TypeScript vision-to-reasoning pipeline with Gemini OCR, Zod-validated model output, deterministic finance calculations and DeepSeek reasoning
• Previously implemented OpenAI function-calling flows in a Node.js/Express SaaS codebase at ByteCroniX
• Portfolio: liuyuelin.dev
• Blog: blog.liuyuelin.dev
• Resume: Yuelin Liu Resume
• Email: liuyuelintop@gmail.com
• LinkedIn: linkedin.com/in/liuyuelintop

• AI/LLM application engineering and AI developer tooling
• Full-stack product engineering with TypeScript, Node.js, Next.js and React
• Reliable AI boundaries with structured outputs, Zod validation and deterministic computation
• Backend/data integration with REST APIs, MongoDB, Convex, pandas and IndexedDB/Dexie
• Engineering confidence through server-side authorisation, automated testing, CI and GitHub-native workflows

A compact, model-neutral contract for AI-assisted software-development work that keeps durable engineering state in Git and GitHub instead of conversation history.

• Defines a GitHub Work Issue contract for the product-discovery-to-execution handoff
• Separates the lifecycle into reusable NEXT → IMPLEMENT → ACCEPT → SHIP Skills
• Ties acceptance to the exact pull-request head SHA and explicit human approval
• Uses repository-native artifacts, deterministic verification and risk-adjusted review rather than parallel status files or context buses

A DeepSeek Harness (DSH) Web plugin that exports the current AI conversation as clean Markdown for reading, Git and cross-assistant handoff.

• Adds an Export Chat action alongside DSH's official Session Log workflow without replacing it
• Exports human-authored messages and final assistant answers while omitting reasoning, tool activity and runtime metadata
• Handles Markdown and Unicode output, incomplete responses, image-only turns and malformed fenced code blocks
• Packaged as an MIT-licensed Node.js plugin; v0.2 was validated with @deepseek-ai/dsh@0.1.0-rc.6

Tech: JavaScript, Node.js, DSH plugin APIs, Markdown

A TypeScript/Node.js vision-to-reasoning pipeline for timecard images and structured wage/finance audits, with explicit data boundaries for local and hosted use.

• Uses Gemini for structured OCR, validates model output with Zod and computes wage/finance metrics deterministically before a DeepSeek reasoning stage
• Separates vision, reasoning and deterministic computation behind provider interfaces, with deterministic mock providers for end-to-end testing without external model calls
• Includes a bearer-protected totals-only extraction endpoint with bounded upload checks and a distinct hosted-data boundary
• Keeps local CLI/library and hosted processing responsibilities explicit rather than making broad privacy guarantees

Tech: TypeScript, Node.js, Gemini, DeepSeek, Zod, structured outputs, deterministic validation

A public Next.js 15 / TypeScript / MongoDB club-management application for rosters, alumni records, events, announcements and video content.

• Hardened server-side authorisation with database-backed user/admin roles and explicit write-path handling
• Turned authorisation and write-path findings into automated regression tests
• GitHub Actions runs typecheck, lint, tests and build on pushes and pull requests

Tech: Next.js 15, TypeScript, MongoDB, Vitest, GitHub Actions

A Python + pandas validation workflow for synthetic client-delivery CSV data.

• Implements six validation rule families with structured ValidationIssue / ValidationResult output and a small CLI
• 12 pytest cases cover required fields, duplicate IDs, domain rules and invalid prices

Tech: Python, pandas, pytest

Languages & Querying:
TypeScript | JavaScript | Python | SQL (coursework)

AI & Validation:
OpenAI function calling | Gemini | DeepSeek | Amazon Bedrock | Zod | Structured model outputs | Deterministic validation boundaries

Frontend:
React 19 | Next.js 15/16 | Tailwind CSS

Backend & Data:
Node.js | Express | REST APIs | MongoDB | Convex | pandas | IndexedDB/Dexie

Authentication & Testing:
Clerk OAuth | NextAuth | Vitest | Playwright | pytest | React Testing Library

Cloud & Delivery:
GitHub Actions | Docker | AWS Lambda | Amazon SQS | Amazon Bedrock | Terraform

Blog | Portfolio