Product judgement
Turn an ambiguous need into a workflow with explicit users, constraints, and success criteria.
AI agent & product engineerMelbourne · Shanghai · Remote
Joye
Huang
I take AI agents from promising demos to products that hold up in real workflows.
Enter the work
I do more than build the interface. I make the AI product work.
From workflow modelling and agent behaviour to full-stack delivery, these are three products I helped move toward real users.
Agent systems
Design clarification, memory, tools, retries, and state recovery as one dependable loop.
Full-stack delivery
Carry the product from interaction and APIs through asynchronous jobs and production behaviour.
01 / 2025—26
Playyy
Brand-aware AI image workspaceA creative environment that keeps brand context, generation, and iteration in one product surface.
- 10+
- AI creation tools
- 4K
- export ready
- 02
- frontier image models
02 / 2025—26
atypica
Commercial research multi-agent systemAt Tezign, I worked on a commercial-research agent workflow spanning clarification, persona interviews, memory, and resilient asynchronous tool execution.
- 300K
- AI personas
- 1M+
- interviews
- <30m
- per research
03 / 2025—26
AIXCut
Agent-assisted video editingI worked across the product and full-stack workflow that helps an editing agent shape scripts, organise source footage, and move toward a deliverable cut.
- 03
- workflow stages
- 01
- continuous agent loop
- FULL
- stack delivery
Turn experience into methods others can use.
I break complex engineering work into talks, diagrams, and public notes—not just to present an answer, but to expose a process that others can test, discuss, and reuse.
Open the full talk ↗- 01 What agent engineering actually contains
Map the real boundary from prompt and context to harnesses, loops, and production behaviour.
- 02 A repeatable loop from interview to production
Make projects, judgement, and failed paths visible so learning becomes testable and repeatable.
- 03 The unglamorous work behind reliable agents
Give repetitive information work to an agent and reserve human attention for judgement and craft.
Writing & Notes
Long-form essays and compact research notes from the things I am building and learning.
Essays05
15 June 2026 From Interviews to Agent Engineering: Why I Stay Calm
↗ 28 May 2026I Grilled Another Agent Candidate: 46-Minute Mock Interview
↗ 23 May 2026Interviews Aren't Exams, They're Mutual Selection
↗ 17 May 2026For Everyone Who Wants Into Agents But Doesn't Know Where
↗ Notes05
24 May 2026 Hermes Agent Memory: A Four-Layer L0–L2 Design
↗ 24 May 2026Prompt Caching in Practice: Cache Design and defer_loading
↗ 20 May 2026Faiss vs Chroma: Vector Store Selection Tradeoffs
↗ 04 May 2026Hermes Agent Memory: Engineering Safety Mechanisms
↗Open source & upstream work
Original projects and contributions to products maintained by other builders.
794+ GitHub Stars Original project · Agent Harness 
01 Learn-Open-Harness
Build an agent harness from first principles: loops, tools, memory, and multi-agent work.
Original project · LLM from scratch 02 minimind-notes
Rebuild a small language model from Transformer through SFT, one experiment at a time.
Upstream contributor · 3 merged / 1 open 03 Token Monitor
Contributed activity heatmaps, fresh provider sync, accurate reset windows, and compact token totals.
07 / OPEN TO OPPORTUNITIES
Let’s put AI into a real product.
I’m open to AI agent and full-stack product engineering opportunities. Ask for my résumé, dig into a project, or tell me what your team is trying to ship.
01 / Team
What is your team building?
02 / Hard part
What is hardest to move forward?
03 / Role
What should this engineer deliver?
See if we fitJOYE / 2026