I was personally invited by the Qwen team to speak at Qwen Meetup Korea, and got to present locally here in Korea yesterday — pretty honored to have been reached out to directly.
The talk was about how I got function calling to work reliably on deeply recursive union types — the stuff the industry generally says doesn't work. With `qwen3-coder-next`, first-try success rate was 6.75%. And the entire Qwen 3.5 model family was hitting 0% on union types due to a consistent double-stringify bug. Both ended up at 100%.
Slides (PPT) are also available in the link — speaker notes are written inside as slide notes if you'd like the full narrative behind each slide.
## TL;DR
1. *AutoBe* — AI backend auto-generation agent. Not text code, but AST data via function calling. 4 AST types + 4-tier compiler validation + self-healing loops.
2. *Typia* — The infrastructure that turns 0% into 100%. A single type automates schema, parser, validator, and feedback generator. Lenient JSON parsing + type coercion + precise validation feedback.
3. *In Praise of Function Calling* — Types eliminate ambiguity. Schemas constrain through absence, not prohibition. Model-neutral, mechanically verifiable, deterministically convergent. Applicable to all engineering domains with validators.
4. *Qwen* — Small models are the best QA engineers. They expose system vulnerabilities large models silently paper over.
5. *6.75% is not failure — it's the first input to the loop.* If you can verify, you converge.
Next subject: function calling for domains where the final answer is unverifiable.
Legal rulings, business decisions: you can't verify the conclusion. But you can verify the process. Model every mandatory reasoning step as typed schema properties, put the verdict last.
CoT becomes 100% structurally guaranteed. Streaming + property order validation catches the AI skipping steps in real time, mid-generation. DeepPartial<T> fills corrections progressively as output streams in.
If you can verify the process, you can harness any domain.
I'd used until 2020, and moved away to Windows laptop.
The key reason of moving is that, Apple suddenly moved Korean/English change to another position. In Korea, the language changing key is positioned at the right of space. And in 2017, Apple suddenly erased the language changing key, and let user to guie "Caps Lock" button instead. And the original "Caps Lock" key, it is working only when pressing the "Caps Lock" key over 2 seconds.
I thought that Apple did a terror to Korean users. However, I just kept using the Macbook just by customizing the language change key to others by the 3rd p
zarty key mapping program.
'
By the way, Apple did more terrible thing at 2020. They suddenly removed "\" (Korean dollar symbol called "won"), and moved it to use the "`" key. The previous "`" key has been disabled when typing Korean. So when writing markdown content, whenever I need codeblock, it makes me crazy.
Furthermore, Mac has critical bug on language changing delay problem. If I press the language change button and keep going typing, the language change is adjusted after the following keys. This is called language chaning delay. Korean users are requesting t fix the bug for 20 years, but never be resolved.
I am not a financial trader, but I had a similar experience while building a chatbot that analyzes the economy and current affairs. I focused on function calls and secured as many papers, news, and statistical APIs as possible (and wrapped them) so that the agent could read them. As the basis for judgment increased, the analysis became more detailed, specific, and sharp.
And I am an open source developer, and I need to make the framework used to build this widely known. And I don't know how to do it either. I am writing on developer communities such as dev.to, but it is not easy.
On the other hand, I am thinking of advertising on Reddit and uploading success stories using the framework to YouTube. If it is a trading bot, wouldn't it be possible to collect success stories or good user experiences, advertise them, film videos, and do interviews?
I think CRUD API may remains (called by chatbot), but CRUD app may not.
I got this insight making a shopping mall chatbot from a backend server. It has 289 API functions, and all of them are properly called in the AI agent.
With this experience, I think that many applications may be transformed to chatbot, and only specific purposes apps will be survived.
I don't think every form UI to CRUD API becomes a chatbot interface. First and foremost is the validation of data that leads to chatbot asking multiple questions to resolve data consistency and correctness. Another way to look at is, with CRUD data, chatbot can fill the form UI if it passes all validation & asking user nod before it updates the system. Not all CRUD will transform to chatbot immediately but they may be augmented by chatbots to make the whole interaction as automated and easier, they'll co-exist until AGI can handle every situation perfect
Hey! If you meant your private repositories, there is a "Allow Access to Private Repos" button in the Import from GitHub modal which can be opened from the left hand side in your dashboard.
The talk was about how I got function calling to work reliably on deeply recursive union types — the stuff the industry generally says doesn't work. With `qwen3-coder-next`, first-try success rate was 6.75%. And the entire Qwen 3.5 model family was hitting 0% on union types due to a consistent double-stringify bug. Both ended up at 100%.
Slides (PPT) are also available in the link — speaker notes are written inside as slide notes if you'd like the full narrative behind each slide.
## TL;DR
1. *AutoBe* — AI backend auto-generation agent. Not text code, but AST data via function calling. 4 AST types + 4-tier compiler validation + self-healing loops. 2. *Typia* — The infrastructure that turns 0% into 100%. A single type automates schema, parser, validator, and feedback generator. Lenient JSON parsing + type coercion + precise validation feedback. 3. *In Praise of Function Calling* — Types eliminate ambiguity. Schemas constrain through absence, not prohibition. Model-neutral, mechanically verifiable, deterministically convergent. Applicable to all engineering domains with validators. 4. *Qwen* — Small models are the best QA engineers. They expose system vulnerabilities large models silently paper over. 5. *6.75% is not failure — it's the first input to the loop.* If you can verify, you converge.