One Designer + AI. Hundreds of Deliverables. — Vincent Wendy, AI Engineer

Read the talk

One Designer, Hundreds of Deliverables: Building a Conference Design Workflow with AI

Selected presentation frame from One Designer + AI. Hundreds of Deliverables. — Vincent Wendy, AI Engineer at 302 seconds
One Designer, Hundreds of Deliverables: Building a Conference Design Workflow with AI

Vincent Wendy explains how explicit design rules, reusable graphics, data-driven exports, and visual checks help one designer support a large conference—and why handling exceptions remains central to the job.

From a talk by Vincent Wendy

At a glance

Ideas worth remembering

  • Explicit typography, colors, and spacing make reusable design possible and give an agent concrete constraints for implementing it.

  • Reusable tools turn repeated work into selection, data loading, and export. Room signage and speaker graphics illustrate that approach, while the flash-drive delivery shows that some operational steps remain manual.

  • Visual matching can reduce photo-search effort and add a check for missing sponsor logos. The reported accuracy reflects the speaker’s experience and tests; the talk supplies no general error rate.

  • Automation must accommodate exceptions. The missing edit button demonstrates why a workflow needs a practical path from a last-minute change to a corrected deliverable.

The scale problem is also an error problem

Selected presentation frame from One Designer + AI. Hundreds of Deliverables. — Vincent Wendy, AI Engineer at 175 seconds
The scale problem is also an error problem

The talk begins as a practical account of designing AI Engineer conferences for an audience made up mostly of engineers. The senior creative designer works in a team of roughly 12 to 15 people, with one designer responsible for a wide range of conference materials. Automation is his answer to the mismatch between that small team and the event’s growing demands.

When he prepared the presentation, the conference expected 6,000 attendees; by the talk, that number had reached 7,000. It also had more than 140 sponsors, more than 300 speakers, and more than 600 sessions. Those counts generate different kinds of work: sponsor assets, speaker graphics, stickers, landing pages, track mascots, physical signs, and digital signage. Scaling design therefore means supporting many formats and many changing records.

The risk grows with the detail count. A missing sponsor logo creates a serious omission, while an incorrect speaker schedule can misdirect people. His observation that a thousand details create a thousand ways to fail establishes the real requirement: a workflow must produce enough assets while preserving completeness and correctness.

0:220:31
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:12 · section reference included

Combine tools around the deliverable

Selected presentation frame from One Designer + AI. Hundreds of Deliverables. — Vincent Wendy, AI Engineer at 197 seconds
Combine tools around the deliverable

He introduces his design team as himself, Devin, GPT, and Figma. His working premise is that available tools have become capable enough for concrete problems to drive the work. Asked what inspires his conference design, he points to the problems themselves: the operational need gives him something specific to solve.

A pelican riding a bicycle illustrates how he works around a capability limit. He describes testing a model’s ability to create that scene directly as a vector file and finding the basic model’s output unusable for his design work. His alternative is to ask ChatGPT for a still image, such as a PNG, and then vectorize it in Figma. The intermediate format changes, but the workflow can still reach a shippable asset. He does not provide a comparison of vector quality or editing effort; the example teaches a way to compose tools rather than a guarantee about conversion fidelity.

3:173:26
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

3:17 · section reference included

Define the design system before generating variations

Selected presentation frame from One Designer + AI. Hundreds of Deliverables. — Vincent Wendy, AI Engineer at 324 seconds
Define the design system before generating variations

His method has five parts: establish the foundation, make designs reusable, automate workflows, validate output, and remove friction. The foundation consists of the design system’s typography, colors, and components. The website embodies those decisions, including primary and accent colors, typography, and the tagline, giving subsequent work an established set of choices.

He connects this approach to his product-design background and atomic design: create the smallest useful parts, then combine them like LEGO pieces into complete deliverables. The mechanism is reuse through composition. A designer establishes the pieces and their relationships once, making it possible to assemble further outputs without deciding every visual detail again.

Those rules also constrain model output. He explicitly supplies desktop and mobile typography because, in his experience, language models otherwise introduce arbitrary font sizes. Defining the values reduces the room for inconsistent choices. The foundation thus serves both as a shared visual language and as a specification the model can follow.

4:524:54
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:52 · section reference included

Reuse the brand and render schedules from current data

Selected presentation frame from One Designer + AI. Hundreds of Deliverables. — Vincent Wendy, AI Engineer at 449 seconds
Reuse the brand and render schedules from current data

Once the website expresses the branding, other teams can use it as the basis for their own materials. He describes marketing creating email designs, flyers, and documents from that reference. Repeated mascot designs follow the same principle: an established visual treatment can be applied again. This distributes production work while retaining a common foundation.

Room schedules provide the first automation example. Previously, he laid them out manually in Figma. The newer workflow uses Devin to pull the latest data for a requested room and set of days, then exports the result as a downloadable PNG. He transfers the image to a flash drive and puts it on the room’s screen. The described pipeline automates data retrieval and graphic production, while display delivery still involves a physical transfer.

He reports that the resulting schedule data is accurate, but the talk does not explain the data source or how accuracy is checked. Pulling current records addresses the burden of manually rebuilding schedules; it does not, by itself, establish that the underlying records are correct.

For visual fidelity, he describes shortening the feedback loop between design and implementation. Earlier handoffs to engineers could require repeated corrections when the result did not match the design precisely. With Devin, he can ask for a more accurate implementation, connect it to MCP, or provide a specification sheet defining spacing and font sizes. The concrete measurements give the agent something more precise to follow than a general request to match the design.

6:486:51
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

6:48 · section reference included

Turn speaker graphics into a reusable product

Selected presentation frame from One Designer + AI. Hundreds of Deliverables. — Vincent Wendy, AI Engineer at 694 seconds
Turn speaker graphics into a reusable product

With more than 300 speakers, producing announcement graphics individually would overwhelm his capacity. He instead demonstrates a speaker-announcement tool with a selection control, editable name text, and a landscape mode. When a speaker has a headshot and the necessary details, he says the tool can export automatically. Trading cards extend the same approach to another deliverable, and he identifies TBPN as the inspiration for one example.

He calls the outputs pixel perfect. His explanation centers on how design information reaches the implementation: Devin lives in Slack, so his working loop moves between Slack, Figma, and Slack again before shipping. He contrasts that direct loop with the more extended research, design, and feedback process he used earlier in his product-design career. He presents this as a change that suits his own work, rather than demonstrating that research is unnecessary for every product.

The specification can arrive through MCP or through a document. He describes a free Figma plugin that adds annotations to a PDF, exposing spacing, font sizes, and colors. Even when layers have generic names such as frame three or frame four, he says the model can interpret the specification. The useful mechanism is the explicit visual measurement; the talk gives no measured tolerance for the claimed pixel-perfect result.

9:179:18
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:17 · section reference included

Find the right speaker photograph

A newer task is finding event photographs for speaker thumbnails. He asks Devin who appears in a photo and describes an interface for deciding whether the pictured person matches a speaker. In the demonstration, he confirms a match as Jason Liu. Previously, thumbnail work involved searching through the photographer’s image codes one by one, sometimes using the time as a clue. A proposed identity now leads him directly to a photo he can download and place in the thumbnail.

He considers the matching fairly accurate, but provides no error rate, evaluation set, or explanation of the recognition method. The demonstrated value is a reduction in search effort with a person checking the proposed match. That experience reinforces his broader enthusiasm: fast design and automation make a worthwhile problem the starting point for useful work. His suggestion that almost anything can then be solved expresses that optimism, rather than establishing a technical capability boundary.

11:4711:49
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

11:47 · section reference included

Use visual comparison to catch missing logos

Selected presentation frame from One Designer + AI. Hundreds of Deliverables. — Vincent Wendy, AI Engineer at 809 seconds
Use visual comparison to catch missing logos

The sponsor banner in the lobby contains logos for more than 140 sponsors. He asks Devin to compare the graphic and check for missing logos, then reports 100% accuracy in the tests he ran. He also uses the same technique on the conference T-shirt. This makes visual inspection part of the production workflow: the agent checks whether a dense graphic contains the required sponsor marks.

The accuracy claim is limited to his own tests. He does not give their number, the comparison reference, or the conditions under which omissions were tested. His practical recommendation is to combine human and AI inspection: a person can overlook a small missing element, and an additional visual check can catch it. The agent supplies another opportunity to detect an error before the design reaches its audience.

13:0713:09
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

13:07 · section reference included

Design the attendee journey and make exceptions manageable

Selected presentation frame from One Designer + AI. Hundreds of Deliverables. — Vincent Wendy, AI Engineer at 928 seconds
Design the attendee journey and make exceptions manageable

Removing friction begins with imagining the experience as a user. He mentally walks through the attendee’s path to registration, wayfinding, QR codes, and rooms. These elements need to connect so people know where to go. The design requirement extends beyond the appearance of individual signs to whether the sequence of information supports navigation.

He then identifies handling exceptions as the real job. The schedule tool initially lacked an edit button. One morning, a schedule needed updating, so he asked Devin to add the control. He reports that it did, allowing him to change the schedule, export a revised PNG, and put it back on the screen. The example shows both a gap in the first tool and the usefulness of being able to extend it when a concrete need appears.

The update still requires exporting and replacing the displayed image. The talk also leaves open how those edits relate to the original schedule data. What he establishes is a shorter path from an unexpected operational requirement to revised signage, which he finds much easier than the earlier manual process.

His closing advice is to solve scale by thinking small: identify the smallest things that can go wrong and try to address them in advance. He ends by returning to the value of having a real problem, because it gives the work a direction and a reason to ship a better product. His broad claim that everything can be automated remains an expression of confidence; the actionable lesson is to focus on specific failure points and make them easier to handle.

14:1714:21
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

14:15 · section reference included

Read the complete timestamped transcript
  1. 0:12

    All right.

  2. 0:14

    Hello everyone. Hope you guys having a

  3. 0:16

    good time at the conference.

  4. 0:18

    So, before we start

  5. 0:20

    how many of you are actually uh

  6. 0:22

    designers? Like a product designer. Hey,

  7. 0:24

    one hands and another. Okay.

  8. 0:28

    And how many I assume that the rest of

  9. 0:31

    you are engineers? Is that correct?

  10. 0:33

    Yeah, pretty much. Okay.

  11. 0:35

    So, today's talk is a non-technical

  12. 0:38

    talk, but more of a real-world

  13. 0:40

    experience how I created the design for

  14. 0:44

    AI Engineer this conference and

  15. 0:46

    our other past conference as well and

  16. 0:48

    how AI has helped me. And so, the talk

  17. 0:53

    today is one designer plus AI, which is

  18. 0:56

    me as the designer,

  19. 0:57

    and hundreds of deliverables.

  20. 1:00

    All right, let's start.

  21. 1:02

    So, my name is Vinson Weng. I am a

  22. 1:04

    senior creative designer at AI Engineer.

  23. 1:07

    And at AI Engineer, it's a very small

  24. 1:10

    team. So, we only have around 12 people

  25. 1:14

    to 15 people at the moment. And

  26. 1:17

    everyone has been doing their own thing

  27. 1:20

    and

  28. 1:21

    I think AI has been like has been a

  29. 1:24

    really helpful way to like helping

  30. 1:27

    everybody doing everything.

  31. 1:29

    And

  32. 1:30

    at if for if an at this scale

  33. 1:33

    we have a problem, obviously, right?

  34. 1:35

    And the problem is the scale problem or

  35. 1:38

    I would call the challenges.

  36. 1:40

    And how to overcome it?

  37. 1:43

    It's basically automation and we get to

  38. 1:46

    that in the later part of this talk.

  39. 1:49

    So,

  40. 1:52

    when I prepared this talk, we only

  41. 1:54

    expected 6,000 attendees and now it's

  42. 1:56

    7,000. Well,

  43. 1:58

    good for us.

  44. 2:00

    And then we have 140 sponsors. More.

  45. 2:04

    140 plus sponsors. And then 300 plus

  46. 2:07

    speakers, 600 plus sessions, and one

  47. 2:09

    designer.

  48. 2:11

    And everybody needs

  49. 2:14

    every designs, right? Like every single

  50. 2:16

    thing needs design. Sponsor needs

  51. 2:18

    assets, speaker needs graphic.

  52. 2:20

    You need sign it so you don't get lost.

  53. 2:23

    And

  54. 2:25

    this is basically what we do, what I do.

  55. 2:28

    So, from stickers, do you like your

  56. 2:30

    swag, your stickers?

  57. 2:32

    Well, I hope you do because I create

  58. 2:34

    that design, too. And

  59. 2:37

    to a landing page,

  60. 2:38

    speaker announcement, track mascot, all

  61. 2:41

    the stuff that you see,

  62. 2:43

    most of the stuff that you see

  63. 2:45

    here, from a sign it to a

  64. 2:48

    digital sign it, landing page,

  65. 2:50

    everything is a deliverable.

  66. 2:52

    And

  67. 2:55

    a thousand details means a thousand way

  68. 2:57

    to fail, right?

  69. 2:58

    Because

  70. 3:01

    I'm missing sponsor logos, going to be a

  71. 3:03

    huge issue. And speakers that have a

  72. 3:06

    wrong schedule, also a huge issues,

  73. 3:08

    right? And it seems impossible to handle

  74. 3:12

    that many kind of deliverables, but

  75. 3:15

    yeah, meet my design team.

  76. 3:17

    So, it's me and Devin, GPT, and Figma.

  77. 3:26

    And right now we are at the stage where

  78. 3:28

    tools isn't the like it's not a problem

  79. 3:31

    anymore, but having a real problem is

  80. 3:33

    our advantage.

  81. 3:34

    So, for example,

  82. 3:37

    when someone asked me, "What inspired

  83. 3:39

    you when designing in AI engineer?"

  84. 3:41

    I don't know the answer back then, but

  85. 3:43

    after I think about it, it's actually a

  86. 3:45

    problem that inspired me to like

  87. 3:47

    designing in this AI engineer. And we'll

  88. 3:50

    get to that in the latter part of this

  89. 3:53

    talk.

  90. 3:54

    So,

  91. 3:56

    have you guys seen the talk by Simon

  92. 3:58

    Wilson like in 2025?

  93. 4:00

    >> Yeah.

  94. 4:02

    >> Yeah, and it's pretty interesting,

  95. 4:03

    right? He asked to

  96. 4:06

    He asked every LLM to create

  97. 4:09

    an a vector file, which is basically a

  98. 4:12

    pelican riding a bicycle.

  99. 4:14

    And it is basically to test and I tested

  100. 4:16

    again and it's still doing this for the

  101. 4:19

    basic model.

  102. 4:21

    And it's not usable for me as a

  103. 4:22

    designer. But as a designer, we have to

  104. 4:25

    think outside the box.

  105. 4:27

    And we could simply ask ChatGPT create a

  106. 4:30

    still image like a PNG for a pelican

  107. 4:32

    riding a bicycle and then I can

  108. 4:33

    vectorize it on on Figma.

  109. 4:36

    And we can ship that now.

  110. 4:38

    So, we have to think outside the box

  111. 4:40

    here

  112. 4:41

    regardless the capabilities of the LLM.

  113. 4:44

    And

  114. 4:46

    So, how to solve this scale problem,

  115. 4:49

    right?

  116. 4:52

    Basically five five things. So,

  117. 4:54

    foundation first, reusable designs,

  118. 4:57

    automated workflows,

  119. 4:58

    validated output, and also remove

  120. 5:00

    frictions.

  121. 5:02

    The foundation is definitely the core

  122. 5:05

    part that we need to set up right. Like

  123. 5:07

    the design system, typography, colors,

  124. 5:09

    components, like other stuff.

  125. 5:12

    And once this is set up, like for

  126. 5:15

    example, when we create the website,

  127. 5:17

    it's all set up within this thing.

  128. 5:20

    And yeah, this is just an example. Like

  129. 5:22

    we have the colors, primary, and then

  130. 5:25

    also the accent colors,

  131. 5:27

    the typography,

  132. 5:29

    and also the tagline, all the other

  133. 5:30

    stuff.

  134. 5:32

    And

  135. 5:34

    also, have you guys Are you guys

  136. 5:35

    familiar with the atomic designs?

  137. 5:38

    So, yeah, my previous background is I'm

  138. 5:40

    a product designer. So, I'm pretty

  139. 5:42

    familiar with the

  140. 5:44

    thing where we need to create a

  141. 5:46

    user-centric design and also like atomic

  142. 5:48

    designs, right? Where we create the

  143. 5:49

    smallest part possible and then

  144. 5:51

    combining it into like basically a LEGO

  145. 5:53

    pieces and then into a deliverables.

  146. 5:57

    And this is pretty useful in my job desk

  147. 6:01

    right now.

  148. 6:02

    So, once we set up all of those

  149. 6:06

    foundation, we basically need to create

  150. 6:09

    for example, we use Defont a lot. In at

  151. 6:11

    the office, we everybody use Defont.

  152. 6:14

    Everybody like

  153. 6:16

    abusing Defont for example.

  154. 6:17

    Yeah. And

  155. 6:20

    So, in this case, I just need hey, we

  156. 6:23

    use this desktop typography and this

  157. 6:25

    mobile typography because we know

  158. 6:29

    Cloud or like any other LLMs love to

  159. 6:33

    like throwing some random

  160. 6:35

    font size, right? And if we don't define

  161. 6:37

    it, it just delivering a slope like the

  162. 6:40

    previous slope.

  163. 6:42

    And

  164. 6:44

    yeah.

  165. 6:45

    Typography, color and stuff and

  166. 6:48

    and then it comes to reusable design.

  167. 6:51

    So, once we set up it right, like the

  168. 6:53

    website is has the

  169. 6:56

    the branding to it, all the other teams

  170. 6:59

    on the AI engineer, like for example,

  171. 7:01

    the marketing teams

  172. 7:03

    can create everything basically. Like

  173. 7:05

    they can create an email design based on

  174. 7:07

    that. They can create a flyer, a

  175. 7:10

    document

  176. 7:11

    just based on the website because it's

  177. 7:13

    already defined

  178. 7:15

    like it defined early.

  179. 7:19

    And yeah, once you get the design, you

  180. 7:21

    can just rinse and repeat.

  181. 7:24

    For example, the mascot, it's all has

  182. 7:26

    the pretty much the same design and it's

  183. 7:28

    rinse and repeat. And if you

  184. 7:33

    already defining those things, you can

  185. 7:34

    basically like create one design that

  186. 7:36

    works for all.

  187. 7:38

    And this is the part that I'm most

  188. 7:39

    interesting to talk about, which is the

  189. 7:42

    automated workflows.

  190. 7:44

    Before, for example,

  191. 7:46

    if you take a look outside the room,

  192. 7:48

    there's a schedule, right? The schedule

  193. 7:50

    for each and everyone.

  194. 7:52

    So, we used to do it manually on Figma,

  195. 7:56

    but now we use Devin for it.

  196. 7:58

    And

  197. 8:00

    let me show you.

  198. 8:04

    Hey.

  199. 8:06

    So, right now we just pull the latest

  200. 8:08

    data. I just asked Devin like, "Hey,

  201. 8:11

    I want this room

  202. 8:13

    at these days." And then we can just

  203. 8:16

    export it, download it PNG, and the data

  204. 8:18

    is all accurate, and then we can just

  205. 8:21

    ship it to the flash drive, and then

  206. 8:24

    put it on the screen.

  207. 8:25

    And

  208. 8:27

    it was like impossible before because

  209. 8:29

    the friction is just too much between

  210. 8:32

    the designers and the developers.

  211. 8:34

    We [snorts] cannot make things like

  212. 8:35

    pixel perfect because

  213. 8:37

    once we tell the designer, "Hey, this is

  214. 8:39

    the design." And then

  215. 8:41

    Sorry, the the engineers that created

  216. 8:43

    the design, for example. "Hey, I need

  217. 8:46

    this to be delivered." And then they

  218. 8:48

    don't create it pixel perfect, it's a

  219. 8:50

    lot of

  220. 8:52

    feedback loop, right? But with Devin, we

  221. 8:54

    just say, "Hey,

  222. 8:57

    can you make this more accurate?"

  223. 8:59

    We can just connect it to MCP, and then

  224. 9:01

    if it doesn't work, we can just always

  225. 9:04

    like

  226. 9:05

    give a spec sheet or something that

  227. 9:08

    can be defined like what's the spacing,

  228. 9:11

    what's the

  229. 9:12

    font size, etc. And

  230. 9:17

    this is what we do to for the speaker

  231. 9:18

    announcement. So,

  232. 9:21

    we have 300 plus speakers, and it's

  233. 9:23

    impossible for me to like handle one by

  234. 9:25

    one, right? So, we create this thing,

  235. 9:28

    which is called which you can also

  236. 9:30

    access to speaker announcement, and you

  237. 9:32

    can also try

  238. 9:34

    it yourself.

  239. 9:36

    Like this one, for example.

  240. 9:38

    You can select it right here.

  241. 9:40

    And then you can also change your name.

  242. 9:42

    Well, that

  243. 9:44

    Yeah. For example, this you can change

  244. 9:46

    the name to whatever you want. And we

  245. 9:49

    also have the landscape mode

  246. 9:51

    which can be also loaded. If the speaker

  247. 9:53

    also have the headshot and all the

  248. 9:56

    details, it will automatically export.

  249. 9:59

    And we also have the trading cards

  250. 10:01

    which is surprisingly pretty popular.

  251. 10:05

    And we have a different team. And this

  252. 10:07

    is all pixel perfect.

  253. 10:10

    All right. For example, this one.

  254. 10:14

    This is inspired by TBPN, so

  255. 10:19

    Yeah. And how do I deliver this in pixel

  256. 10:22

    perfect? Let's jump into it.

  257. 10:25

    So,

  258. 10:27

    the process here is

  259. 10:31

    before

  260. 10:32

    when I start my career as a product

  261. 10:35

    designer, it used to be just

  262. 10:37

    okay, we need to research, we need to

  263. 10:40

    build product like design thinking in

  264. 10:41

    general, right? And then feedback loop

  265. 10:44

    and stuff like that. But right now, it's

  266. 10:47

    it's just outdated for me. Like in my

  267. 10:50

    case,

  268. 10:51

    we just go to Slack,

  269. 10:53

    Figma, and then send it back to Slack

  270. 10:55

    because Devin or Devin live in Slack,

  271. 10:58

    and then ship all the things that he

  272. 11:00

    need.

  273. 11:01

    >> [snorts]

  274. 11:01

    >> Like for example, if we can connect the

  275. 11:04

    MCP or also the

  276. 11:06

    spec document, which is for example the

  277. 11:09

    spec sheet like this,

  278. 11:10

    which is

  279. 11:11

    uh plugin in Figma if you interested.

  280. 11:15

    It's free and it's basically give an

  281. 11:17

    annotation to the PDF.

  282. 11:20

    And

  283. 11:22

    yeah, all designers don't name their

  284. 11:24

    layers, so yeah, this is just like

  285. 11:27

    some random frame three, frame four, but

  286. 11:29

    the LLM will get it.

  287. 11:31

    And it's basically defining all this

  288. 11:34

    spacing, all this

  289. 11:36

    font size, and then all the

  290. 11:39

    colors and stuff. It's definitely going

  291. 11:40

    to help you develop a pixel-perfect

  292. 11:43

    product.

  293. 11:44

    And

  294. 11:47

    we also have just recently like today

  295. 11:49

    have uh photos which we have to create

  296. 11:52

    the thumbnail for its

  297. 11:54

    speaker, right? And then we ask Devin

  298. 11:56

    like, "Hey, who is this person?" And

  299. 12:00

    yeah, it kind of did. Like I make a

  300. 12:03

    Tinder kind of

  301. 12:05

    you know, detection

  302. 12:07

    if this is the same person or not. And I

  303. 12:09

    think it's pretty accurate.

  304. 12:11

    It's Jason Liu. Yes.

  305. 12:13

    And then we can use this to like

  306. 12:17

    for context. Like before, when we create

  307. 12:19

    the thumbnail, we have to search all the

  308. 12:21

    codes that photographer have and search

  309. 12:24

    it one by one and maybe by time

  310. 12:26

    if possible. But now we can just like,

  311. 12:29

    "Oh, this is Jason Liu. Download that

  312. 12:31

    photo." And then we can paste it into

  313. 12:32

    the thumbnail, right? And it's pretty

  314. 12:35

    amazing. I mean, the world that we live

  315. 12:38

    in right now is actually like the state

  316. 12:41

    for me as a designer is already at the

  317. 12:43

    peak because

  318. 12:45

    what what else can you ask for, right? I

  319. 12:48

    mean, we already have things to

  320. 12:49

    automate, we already have things to

  321. 12:51

    create the design fast.

  322. 12:53

    Basically, all you need is a problem

  323. 12:56

    because once you have a problem that

  324. 12:57

    worth solving, you can

  325. 13:00

    basically solve anything.

  326. 13:02

    And back to my talk, I got sidetracked

  327. 13:05

    right there.

  328. 13:06

    Yeah.

  329. 13:07

    And then yeah. And this is also the

  330. 13:09

    amazing thing that we test.

  331. 13:11

    So, as you know, we have like hundreds

  332. 13:15

    of sponsors, right? Like 140 plus. And

  333. 13:18

    as you can see on the at the lobby, we

  334. 13:21

    have the banner with all the sponsors.

  335. 13:24

    And

  336. 13:25

    I basically tell Devin like

  337. 13:27

    "Hi, could you compare

  338. 13:29

    could you check if there are any missing

  339. 13:31

    logos in this graphic?" And the accuracy

  340. 13:34

    is 100% based on the test that I

  341. 13:38

    do. So,

  342. 13:40

    which is pretty well. And we use the

  343. 13:41

    same thing for the

  344. 13:45

    T-shirt that you got for your swag.

  345. 13:48

    And yeah, surprisingly, Devin knows how

  346. 13:52

    to like visualize things, right? Like

  347. 13:54

    how to detect things visually.

  348. 13:57

    And that is very surprising because

  349. 13:59

    as a human, we can like give errors. Oh,

  350. 14:03

    turns out there's one small something

  351. 14:04

    that is missing. But with this kind of

  352. 14:07

    thing, we can like double-check. So,

  353. 14:09

    human plus AI, combine it,

  354. 14:13

    well, you got your own QA team.

  355. 14:15

    And then remove fiction.

  356. 14:17

    So, this is just uh the way of thinking.

  357. 14:21

    So, as a designer, we have to think

  358. 14:24

    as a user, not as a designer, all right?

  359. 14:27

    Because every user has its needs.

  360. 14:29

    You can walk through the for example,

  361. 14:31

    the map plan here. So, basically,

  362. 14:35

    I'm imagining myself as an attendee to

  363. 14:37

    go to the registration, go to the

  364. 14:41

    see the wayfinding and the QR code and

  365. 14:44

    then all the stuff. Basically,

  366. 14:46

    everything needs to be connected so you

  367. 14:48

    guys don't get lost and knows how to

  368. 14:51

    find your rooms and

  369. 14:54

    other stuff.

  370. 14:55

    And

  371. 14:57

    the real job is handling exceptions.

  372. 14:59

    So, for example, oh, I have Yeah.

  373. 15:03

    >> [snorts]

  374. 15:03

    >> For example,

  375. 15:04

    um

  376. 15:06

    there is a schedule update, all right?

  377. 15:09

    And

  378. 15:10

    when we create this thing, it doesn't

  379. 15:12

    has an edit button. And then one

  380. 15:15

    morning, it just "Hey,

  381. 15:17

    this schedule needs to be updated and we

  382. 15:19

    don't have those edit buttons." I could

  383. 15:21

    just ask Devin, "Hey, can you add me an

  384. 15:23

    edit button?" And then it did. So, we

  385. 15:26

    can change everything now and then ship

  386. 15:28

    it to PNG and replug it to the screen,

  387. 15:32

    which is pretty convenient, right? And

  388. 15:34

    those exceptions, right? It

  389. 15:37

    it's not possible before

  390. 15:39

    when we have to do it manually and

  391. 15:40

    stuff.

  392. 15:41

    But now it's just get easier. And

  393. 15:46

    so, the takeaway here is that to solve

  394. 15:49

    the scale problem, you have to actually

  395. 15:51

    think small. Think all the smallest

  396. 15:52

    thing possible. Think everything that

  397. 15:55

    can go wrong and will go wrong and then

  398. 15:57

    try to solve it before. And also, like

  399. 16:02

    yeah, right now basically you can

  400. 16:03

    automate everything.

  401. 16:05

    And

  402. 16:08

    at this moment, having a problem is

  403. 16:10

    actually going to benefit you because

  404. 16:12

    that's going to help you ship

  405. 16:14

    a better product, going to ship uh

  406. 16:16

    things that are

  407. 16:17

    good. And yeah, I think that's all that

  408. 16:20

    I can share. Hope my talk has some

  409. 16:23

    benefits to you and

  410. 16:24

    yeah.

  411. 16:25

    That's all. Thanks, guys.

  412. 16:27

    >> [applause]

  413. 16:46

    [music]