Caesar Chi：當開發遇上包裝：AI如何讓產品從想法變成商品

1. In one sentence

The enormous gap in situational information between a brand's creative end and a factory's manufacturing end is what this session narrowed down step by step, using an intent detector, a supervisor architecture and human-in-the-loop, and validating it repeatedly through physical sampling.

2. The session

The speaker, Caesar Chi, Chinese name 戚務漢, works at Cympotek (信集界科技), part of the Cymmetrik Group, where he leads development of Cympack, an AI and manufacturing integration platform. He is also a core member of the JSDC team, a long-time driver of Taiwan's Node.js community, a co-founder of JSDC.tw, and the founder of EXMA-Square (verified against his public speaker page and Speaker Deck profile, source one, source two). The host introduced him on stage as having co-founded an "AI for Developer" community with friends in 2023; the exact relationship between that community name and EXMA-Square is unverified.

Company background: Cympotek and Cympack

• Cympotek (信集界科技) is part of the Cymmetrik Group, whose parent is the largest professional packaging and label printing service provider in Greater China (verified against the Cymmetrik site, confirming Cympotek is one of its subsidiaries, source). Cympotek itself focuses on brand packaging materials, new retail and unique-code-per-item solutions, aiming to build a high-stickiness user experience between brands and consumers.
• He mentioned the team is around fifteen people, covering micro-factory operations, development, procurement, administration and pre-press handling — small and nimble (speaker's own account, no external source). The company figures he gave: fifteen locations worldwide, over fifty-five years of experience, Fortune 500 international clients, over thirty selected global supply chain partners, and over ten international clients (speaker's own account, no external source).
• What the team does, he calls a "manufacturing AI support chain": letting brands complete a new product introduction or new packaging delivery through AI. They call the whole system Cympack.

From idea to delivery: four stages, and a coincidence with the "second brain"

• His observation is that developing a new product or new packaging falls into roughly four stages: collection → analytics → prototype → product. The collection stage iterates continuously, turning chaotic data into useful information; only then do you move into market analysis, prototyping and finally the product.
• He deliberately stressed running the whole flow, to avoid "pulling something out of thin air and producing a mess". A lot of products end up inexplicable, in his view, precisely because iteration and analysis were skipped.
• The team found these four stages coincidentally align with the CODE model from the *Building a Second Brain* methodology (Capture, Organize, Distill, Express); both are a process of going from small to large and then back from large to small.
• At first the team assumed the flow was linear, so they used RAG for the data wrangling and put together an OpenAI/LLM-backed API wrapper to solve each problem. He described this stage as "looking like everything was blissful" — but it was nothing of the sort, because the collection stage itself requires extensive detailed discussion, data definition, execution confirmation and regression on issues, and cannot be solved off the top of your head.

A case: the semantic gap between a paper box and a packaging box

• His example was "make me a paper box": given only the word "paper box", that could mean at least five completely different processes — hand cutting and folding, die-cut creasing, laser cutting, smart cutting, or a digital proofing machine at a print house.
• But what the client actually wants is usually a "packaging box", not a "paper box" — and he said the manufacturing processes and techniques for the two are completely different. The client simply did not know which word described what they wanted, and set off in the wrong direction from the start. He was careful to add: "this doesn't mean it's the client's fault" — often the two sides are just not on the same frequency, and it is hard to say whose problem it is.
• The critical piece the team had been overlooking is called "the scenario": what situation the user actually intends to use it in. So they added an intent detector to the architecture, continuously extracting the client's real intent from the conversation to keep communication quality consistent on both sides.
• After bringing in AI agents, the team decomposed the flow between analysis and prototyping into finer task units and switched to a supervisor architecture to orchestrate the whole agent flow: the user supplies information or context, and the supervisor decides which agent should handle it.
• The whole flow finally simplifies to: data straight to information, then into sampling and prototyping — a loop that may repeat several times — and only then converging on the product.

Sampling: the key to a product being born, and you do it several times

• His words: "The key factors in a product being born: first, sampling; second, sampling; third, still sampling."
• A case: a client brought a simple front-facing image of a figurine and asked whether it could be made into a plush toy. The team assessed it, said yes, and with AI analysis plus a cooperating supplier actually produced a sample the client was very happy with — they nearly placed an order for twenty thousand on the spot. The team suggested sampling a few more times, and it turned out the hat had not been glued properly, so taking it off left the toy completely bald; and because the client had only supplied a front view, the team had never designed the back at all, making the result "completely unsellable". He called this "the most successful failure", and stressed it happens every day.
• The conclusion: sample at least once before shipping, and keep sampling continuously and in parallel, learning from cases about the client's habits, so you avoid the unrecoverable waste of shipping something that cannot be fixed.

Human-in-the-loop: leave the key decisions to the client

• The team also stores users' past experience and preferences, so that next time the system serves the same client it understands better what they want and which attributes apply, avoiding a repeat of the same problem.
• His observation: a fully automated flow sometimes satisfies the user and sometimes leaves them feeling let down, so the team decided to give the client back a say in important decisions, integrating CopilotKit's human-in-the-loop mechanism. As he described it, that mechanism has two modes, interrupt and node-based, letting users approve/reject, review and edit state, and review tool calls, with the interface never blocking the task but instead entering a waiting state.
• His analogy: the history of the F1 gearbox — F1 ultimately chose neither fully automatic nor purely manual but a semi-automatic sequential gearbox, with extremely fast shifts (he mentioned about 0.02 seconds), improving lap times and stability while preserving room for the driver's skill and strengthening safety. That, he said, is the right proportion for human-machine collaboration: "I know AI really is better than me sometimes, but I still want to decide some things myself."

Architecture evolution: CrewAI and the next stage of multi-agent orchestration

• As the number of nodes in the flow grew — document parsing, pre-press checks, model analysis, production work order generation, scheduling optimisation, data reporting, customer service feedback — the team found the architecture needed continuous improvement to reach a genuinely fully automated delivery flow on the factory side.
• The team later migrated to CrewAI, adopting four concepts: Crew (the top-level organisation, managing the AI agent team, supervising the workflow, ensuring collaboration and delivering results), AI Agents (specialised members with roles such as researcher or manufacturer, able to use designated tools, delegate tasks, make autonomous decisions and share state), Tasks (individual jobs with clear goals and tools, producing usable output) and Process (flow management, defining the collaboration pattern and controlling task assignment). He said this architecture greatly simplified the code and the overall structure (verified against CrewAI's documentation; the four concepts of Crew / Agents / Tasks / Process match his description, source). The team then combined it with LangChain, heading toward the next stage of agent flows.

Closing: how to collaborate in the post-AI era

• In his view, the most important part of the whole system is delivery: how to turn chaotic data into data a machine can read, and finally into a physical sample and a product handed to the client.
• On how to collaborate with AI in the post-AI era, his metaphor is skiing: you cannot forcibly drag AI onto the exact line you wanted to take; you go with the flow. What matters is consciously accepting that AI is imperfect and focusing on the goal itself, moving toward it.
• His closing line, kept in his own words: "The post-AI era is really about accepting that feeling of being slightly out of control and partly in control at the same time."

3. Figures and cases

• A team of about fifteen, covering micro-factory operations, development, procurement, administration and pre-press (speaker's own account, no external source).
• Company figures as stated: fifteen locations worldwide, over fifty-five years of experience, Fortune 500 international clients, over thirty selected global supply chain partners, over ten international clients (speaker's own account, no external source).
• The contrast: "paper box" maps to five processes (hand cutting and folding, die-cut creasing, laser cutting, smart cutting, print-house digital proofing machine), "packaging box" to five different ones (hand box-making, die-cut creasing, digital laser cutting, paper-mounted forming, automatic folder-gluer) — one requirement, different words, entirely different processes.
• The figurine sampling failure: the hat was not glued properly, leaving the toy bald; designing only from a front view meant the back was never made; verdict, "completely unsellable".
• The F1 semi-automatic sequential gearbox shifts in about 0.02 seconds (speaker's own account, no external source).
• Technical evolution: RAG + LLM API wrapper → intent detector + supervisor agent architecture → integrating CopilotKit's human-in-the-loop → migrating to CrewAI, then combining with LangChain.

4. Lines worth keeping

• "The key factors in a product being born: first, sampling; second, sampling; third, still sampling."
• "This is the most successful failure."
• "I know AI really is better than me sometimes, but I still want to decide some things myself."
• "The post-AI era is really about accepting that feeling of being slightly out of control and partly in control at the same time."

5. Tools and terms mentioned

Manufacturing AI support chain, new product introduction (NPI), the CODE methodology (Capture / Organize / Distill / Express, from *Building a Second Brain*), RAG, OpenAI/LLM API wrapper, intent detector, supervisor architecture, agent flow, CopilotKit, human-in-the-loop (interrupt and node-based modes), CrewAI (Crew / AI Agents / Tasks / Process), LangChain, the Cympack platform, Cympotek (信集界科技), the Cymmetrik Group (正美集團).

6. Wider observations

The most striking part is the paper box versus packaging box case: one requirement, one wrong word from the client, and the entire manufacturing process changes. That kind of situational information gap happens daily in software development too (the product manager's "faster" and the engineer's "faster" are often not the same thing), except that in physical manufacturing the cost of the gap is twenty thousand ruined packaging boxes rather than one refactor. Using an intent detector to keep extracting the real intent is fundamentally the same as software engineering's insistence on clarifying the requirement before you start — just packaged as an agent architecture.

"The key factor in a product being born is sampling, sampling and still sampling" maps directly onto software testing: sampling more times is testing more times, the difference being that a software test costs approximately nothing while a physical sample can mean real money in tooling or materials. That also explains why he made "sample at least once before shipping" a hard rule in Cympack rather than an option.

More surprising still is that the team's path from RAG plus API wrapper all the way to CrewAI and LangChain is almost identical to the evolution of a pure-software agent system — except that what sits at the far end is not a database or an API but pre-press checks and tooling schedules. That suggests the "AI agent architecture" methodology has already begun to seep beyond software into settings that have to deal with physical supply chains, rather than staying inside chatbots and coding assistants. And that closing line about accepting the feeling of being slightly out of control and partly in control applies to any team adopting AI, not just the packaging industry.

7. Sources