Anthropic × Bloom: Ambition for Korea

The Ambition for Korea hackathon, held Saturday, October 24, at Seoul AI Hub, wrapped up successfully.

About 100 builders, picked at a 10-to-1 ratio after we carefully reviewed their GitHub, LinkedIn, and applications, gathered in one place. Of about 40 teams (2 to 4 people each), 8 made the final, and 5 of those matched with VCs.

First of all, thank you sincerely to everyone who showed interest early and applied. It was a real shame we couldn't host everyone. True to an AI hackathon, we ran the whole thing, planning and operations, with AI doing real-time checks, status updates, and feedback loops, and we learned a lot. We'll use that experience to come back bigger next time. I designed the hackathon program to be as tense as Show Me the Money (a Korean rap competition show), and so you can see what the atmosphere was like, I'm open-sourcing the hackathon program that Claude and I built together.

There are so many people to thank. Special thanks to Yeop Lee at Anthropic, who planned the theme of this hackathon with us and discussed it in depth. Giving every participant $500 in credits was already generous, and then, as a surprise on the day, Anthropic sponsored an extra $5,000 in credits for each of the 8 finalist teams, which pushed everyone's focus even higher. Thank you for support worth more than 100 million won (about $70,000). It was great to hear, through the fireside chat, Anthropic's vision for the Korean market and concrete talk about supporting startups as they expand overseas and grow. Bloom started in April this year with the Claude Community under the banner "Claude Blue to Bloom," so this partnership feels especially meaningful.

Deep thanks also to Taeyang Lee of BASS Ventures, Seokwon Hong of AttentionX, Jaewon Lee of SCHMIDT, Kyungpyo Ko of SpringCamp, Taehong Park of SV Investment, and Shinhye Lee of Goodwater Capital, who carefully reviewed every team's demo link and deck PDF from the afternoon, once building was underway, until late at night. Looking back, it was grueling work, judging a team roughly every five minutes, and yet they still left detailed one-line reviews that gave many participants real insight and inspiration. In the evening session they judged team pitches Shark Tank style and went on to give generous feedback on the spot and set up follow-up meetings. I bow my head once more in thanks for that passion.

Above all, deep thanks to Hyunhee Shim of Seoul AI Hub and the Seoul Metropolitan Government for supporting us so we could hold the event in such a great space. Thank you to CJ CheilJedang and Snowflake for sponsoring delicious snacks and drinks. And as always, thank you to Junsu Kim, Yeil Kim, Jaemin Yu, Jeongmin Kang, and Jonghyuk Lee, who worked behind the scenes to keep things running smoothly.

Thank you again to everyone who joined the Ambition for Korea hackathon. I hope this wasn't a one-off event but a Saturday that becomes a first step, and a benchmark, for Korean builders reaching out to the world. We'll keep coming back in better forms and keep sharing behind-the-scenes stories and insights.

Today is Day Zero

In the opening, I started with why we were holding this hackathon. Bloom wasn't built from a strategy or a plan. It came about naturally in answer to the call of the times. When we held our first event with the Claude Community in April, 2,100 people applied in three days. We couldn't host them all at once, so we kept saying "see you next time, see you next time," and that's how we've kept events going for five months now. In that time, 8,000 people have joined more than 40 events, and in July we also held an event with the Singapore Claude Community, so I think of Bloom as having grown up alongside Claude from the start. That made this day, when we officially teamed up with Anthropic, a very meaningful one for Bloom.

Since we had a chance to work together, Yeop Lee at Anthropic and I talked freely about what kind of event to do, and the keyword that came out of it was Ambition for Korea. Yeop leads startup events and programs across Asia-Pacific, including Australia, India, and Japan, but, as a Korean, wanted to cheer on Korean companies going out into bigger markets. Anthropic was taking the Korean market seriously enough to open a Seoul office this June. Once we decided to hold our first one-day hackathon together, the Seoul Metropolitan Government, Seoul AI Hub, and six VCs who shared the goal joined one after another.

"Going global" has become a cliché. But because we all know how hard it actually is, I always felt a little embarrassed saying it. Were we not serious, or were we really serious? So I called the day Day Zero and asked everyone to form teams and take a real shot at it, very seriously, even if only for a day. One day won't settle anyone's global expansion, but I prepared it hoping people would use this chance as a foothold and leave with the drive to keep going.

There are many reasons going global is hard: business, operations, funding, technology partnerships. But I said, "Korea's a pretty big market, you know. So I think there's a bit of a trap in that." For this one day, I wanted to break the usual sequence of trying it first in Seoul, in Korea, and only then going global. These were new teams anyway, so I asked them to break from whatever they'd been doing and treat today as Day Zero, rebuilding their business to target overseas markets from the start.

I also introduced who was with us. Anthropic covered technology. On the government side, Seoul AI Hub, which gave us the space, and the Seoul Metropolitan Government. On the business side, six VCs we chose from among those that help companies go global or run global funds. We also got sponsors for food to get through the day. CJ CheilJedang's bibigo sent 120 each of Hetbahn Faro nurungji tea (a roasted rice tea) and bibigo seaweed nachos, and Snowflake sent 120 iced Americanos with cheering stickers on them. Snowflake's coffee was a prize Bloom won by taking first place in a competition we'd entered while working with Snowflake. I said it would be fun to hear how each VC's vision and direction differ, and slipped in that we'd talk today about whether there might be support beyond credits and try to get a surprise.

Wooseok Byeon, director of Seoul AI Hub, opened with, "My name is Byeon Wooseok, and I'm sorry I look like this," a nod to the famous actor with the same name. Encouraging the office workers who came out on a Saturday they'd rather sleep in, the director said, "I sincerely hope your small attempts today become a first step toward founding companies and even going global." Seoul AI Hub, an agency under the Seoul city government, runs an industrial AX (AI transformation) program supporting open innovation where large and mid-sized companies grow together with startups, acceleration for 100 AI-based resident startups, a membership program open to anyone with about 400 member companies, free education, and overseas expansion programs. Next year it plans to grow the overseas programs, set up a global big tech center inside the hub with a space and programs where people can meet Anthropic regularly, and work with other big tech companies too.

Next, Jieun Park, team lead in the AI Strategic Industries Division of the Seoul city government's Economic Policy Office, said the passion of people who came out on a Saturday morning was rubbing off on the team too, and shared some news. Tortoise Media in the UK this year ranked AI industry capability by city for the first time, and Seoul came in third in the world. In the country rankings, Korea had been fifth. Park credited that result to the people in the room.

I think this result reflects the efforts of people like you, who come out on a Saturday of your own accord to build AI technology.

Promising that the city would strongly support events the hub holds in the future, Park said it was a shame not to be able to stay for the VC judging in the evening, and left.

Building got easy. Selling is the hard part.

I moderated the fireside chat with Yeop Lee, Head of APAC Startup Platforms at Anthropic. Yeop builds and runs Anthropic's startup programs across Asia-Pacific, including the credit programs, and started by saying Anthropic runs a huge number of hackathons around the world, but the energy in Korea is unusually hot. These days Anthropic mostly runs short hackathons where building has to finish within two hours, Yeop said, adding that they've entered about four hackathons themselves and never won a prize, so events like this are exciting.

Yeop joined Anthropic last September as its first team member in Korea, sees Korea as a very important market and ecosystem, and gave some numbers. Claude usage per capita in Korea is about four times the global average. Globally, automation, handing off work entirely, and augmentation, AI helping with the work, are close to half and half, but in Korea augmentation is higher. Because Korea thinks that deeply about how to use AI, feedback from Korean startups has often gone to the Claude Code team and the model team and been reflected.

Asked what sets Korean teams apart, Yeop pointed to mindset.

It's a strength, and also something that would be good to shore up a bit. There seems to be a mindset of testing in Korea first and then going global.

That could mean preparing well in Korea and building competitiveness before heading out. But Indian teams, for example, often start from wanting to take what they built in India straight to the US or go after Europe, so the strategies differ, Yeop explained. Part of why Anthropic co-planned this hackathon is that the more teams go global from day one, the more Anthropic can help. That was exactly the point I'd made in the opening about aiming global from the start.

Anthropic's first Asia-Pacific office opened last October, and in a year there are five. The fifth, in Singapore, opened just last week. When I asked about the pros and cons of covering Asia-Pacific from Korea, Yeop said ending up in Korea was actually pretty much chance, and explained that Anthropic doesn't have a separate Asia-Pacific hub. Yeop is in Korea, the APAC finance lead is in Japan, the applied AI technical lead is in Australia, and so on. The idea is to move quickly and flatly instead of building the layers of hierarchy you get when everyone's in one place. Everyone belongs to one team at headquarters, not to a local branch, so feedback from Korean startups and leaders doesn't stay within the Korea team. It's shared directly with headquarters and teams in other countries.

On this event's setup, where VCs, government, and a community put it together, Yeop said it's far more effective than Anthropic going it alone, because together they can give builders resources to grow faster. I agreed. Hackathons used to take three days and two nights, and building was the fun part, but now that AI produces a prototype in no time, people even say hackathons have gotten boring. So the idea behind this event was to tie everything together in one day, from Anthropic's technical support to VCs checking business viability, so that ideas don't end in the room but lead to a business.

Asked what kind of teams they hope come out of Korea, Yeop said teams that solve problems at a global scale. Anthropic does a lot of B2B and works with a huge number of companies, so these days it also runs a program that connects startups directly with enterprise customers to help with sales, and it's expanding its baseline credit benefits beyond API credits.

Since Bloom started from Claude Blue, I also asked whether these days feel closer to Blue or to Bloom. Yeop said it happens every time a new model comes out and told their own story. Yeop gets hundreds of emails a day and has a personal goal of keeping unread mail under 30. Now Claude writes about 90% of their emails, pulling in Yeop's tone and real materials, and Yeop only checks and sends. Getting there took a lot of trial and error.

Everyone has a realistic worry that if you keep using AI, your ability to think critically on your own will weaken. Yeop talked about this with Anthropic co-founder Ben Mann and found that Ben, too, sets aside dedicated time for structured reading and deliberately does some things himself even when AI is much better at them, to keep a balance. That was Yeop's worry and answer. I added that since there's no single right answer, we should share and figure it out together. Claude also has a mode that asks questions and guides your thinking instead of giving the answer right away, and I read that as a sign it was designed from the start to let users stay in charge of their own thinking while they use AI.

Asked about the mood among builders abroad, Yeop brought up teenage founders. You rarely see a 14-year-old founder in Korea, but abroad there are more and more. The keynote speaker at Australia's biggest startup festival, which Anthropic sponsored in a big way, was 14. That speaker couldn't come alone and came with their parents. Technology opens the door, and across the ecosystem there's a sense that VCs and government should support young founders too. I added a story about a teenager working with my friends in San Francisco who dropped out of high school and still wrote several papers by recruiting people on X, and that a YouTube video we recently shot at Bloom with a 17-year-old high school student got 300,000 views and close to 400 comments. It made me feel how much the walls of credentials, age, and experience have come down in the AI era.

Finally, I asked what matters more than technology, credits, and tokens. Yeop said go-to-market. Building has already gotten very easy, and so has building a convincing product. Big companies have started building their own tools with AI, and the quality of what comes out internally is pretty high. So understanding the end customer's problem precisely matters as much as the skill to build a good product fast, and the most common failure Yeop has seen meeting builders comes from that.

Building is so fun and easy that people just build. They launch, and if the response isn't great, they say it's because it's missing this feature, and with not enough strategic thinking, they go back to building.

That gives you the illusion of progress, but a year later sales are often weak. So the advice was to think about whether you're properly validating the hypotheses you set, and whether you can speed up that validation. I said that people say these days that distribution is the moat, but distribution isn't easy either, and to know if your product is good, showing it to friends isn't enough. Meeting people from other fields at gatherings like this and checking their reactions is part of that validation, and that's also why we brought VCs in as judges.

The surprise I'd hinted at in the opening came at the end of the fireside chat. Yeop decided to give each team that made the final an extra $5,000 in Claude credits. It went to each team regardless of size, to split within the team.

Selection to deadline: how we designed the day

The day ran 12 hours, opening at 9 a.m. and closing at 9 p.m. At 9:00, opening remarks from the Seoul city government. At 9:20, the Anthropic fireside chat. At 9:40, the mission reveal and credit info. Then seven hours of building from 10 a.m. to 5 p.m., with VC judging running in the order submissions came in. Submissions closed at 5, everyone ate dinner while we prepared Shark Tank, and from 7 p.m. the 8 finalist teams each took the stage for 10 minutes, followed by VC and team matching results from 8:30. The opening and Shark Tank were on the 4th floor, which has a stage and an LED wall, and building happened on the 3rd and 4th floors.

We took applications per person, not per team. Even if you came as a team, every member had to apply separately. Anyone 18 or older could apply for free, and the form asked for one link to your own work and a sentence or two on what you wanted to build for an overseas market. 960 people applied, and after going through GitHub, LinkedIn, and applications one by one, we invited 105. Teams were 2 to 4 people, 41 teams in all: 24 two-person teams (the most common), 11 three-person teams, and 6 four-person teams. The opening slide rounded this to 100 out of 1,000, 10 to 1, and made clear from the start that only 8 teams would make Shark Tank.

On the participant page, entering the email you applied with brought up your team. Some people used a different email to get credits, so we linked them in advance so any of their emails would find the same team. Around lunch we reminded everyone that the Shark Tank pitch would be in English, but decks and submission forms could be in English or Korean, whichever was easier.

Instead of only asking for a two-slide deck, we handed out a kit in advance. Most participants would probably build slides with AI like Claude rather than open PowerPoint, so we bundled a guide doc you could hand straight to AI, an HTML template, logo strips for light and dark slides, and partner logos. Six of the eight finalists put the logo strip in as is, and the Shark Tank LED showed each team's deck inside the event frame with the logo strip at the bottom. The idea was that wherever anyone took a photo from, the team's screen, Ambition for Korea, and the partner logos would all land in one shot.

Same $500, every repo public

The rules started from making everyone compete with the same resources. Technology is the same for everyone, so the only thing that makes a difference is the idea and the execution. Each participant got $500 in Claude credits, and that was it. Those credits covered both Claude Code and the API for the deployed app, and teammates could pool them in one organization. We could have given unlimited credits but deliberately set a cap, and we didn't allow topping up with personal or company accounts. Core AI functions like judgment and generation had to be built only with Claude. Peripheral tools for design or video were fine, but at your own cost.

There's no perfect way to verify a rule that only what's built today gets judged. So we revealed the missions only that morning, and made everything else tedious to cheat and easy to catch. Planning and development had to start that day, on site. Code could only go in a new public repo created after 10 a.m., and every team's commit history was public.

The site read each team's declared repo's public commits every few minutes and doubled as a warning system. If the first commit was before 10 a.m., it flagged "Commit before 10." If the first commit alone brought in more than 3,000 lines of code, it flagged "Large first commit." These orange tags appeared next to team names in the ops console and on the VC judging screen. There was also a flag for repos that already had more than 20 commits the first time we read them, but not a single team got it. When Claude proposed this system around midday the day before, I asked that it explain to the VCs why it was being shown. So clicking "Why look at commit history?" on the judging screen opened an explanation that we'd agreed to judge only what was built after 10 a.m. today, along with a note: "Even starting from a template, a repo can get big, so please use this only as a reference for your judgment."

The flags weren't shown on participants' screens, and a flag didn't automatically disqualify anyone. The reference panel on the judging screen showed what the team said they built today, the number of commits, and the time of the first commit, with a GitHub link straight to the code at the moment of submission. If a VC saw a flag, it was up to them to open the history and decide. Of the 36 teams that submitted, 4 got "Commit before 10" and 3 got "Large first commit."

That morning we added that participants could use public open source, but shouldn't fork it. Bring it into a new repo and credit the source, because forking drags in the original project's old commits and triggers the "Commit before 10" flag. The rules also said never commit API keys (hide them with .env and .gitignore), and keep the app running until 9 p.m. so VCs could try the demo themselves. We locked each submission to the commit at the moment it was submitted, so nobody could change the code afterward to change the judging. I knew some people might not love making all their code public, but in the opening I said this:

I think ideas are a public good these days. What matters is how far you push an idea, so I'd love it if you shared your ideas openly, with the attitude that you'll just keep moving ahead.

I meant: instead of kicking away the ladder in case someone follows you, don't worry about it and keep moving forward.

Submission happened in two steps on the same page. First, by 10:30 a.m., one person per team declared a topic: one mission, a one-line idea, the target overseas market, and a public GitHub repo. The declared topic could be edited only until the deadline. The final work could be submitted only once per team, by 5 p.m. It asked for the service name, a one-line intro, the target customer, a demo link, demo instructions telling VCs what to click first, a deck PDF of no more than two slides, plus a separate note on what was newly built today. There was a box to check if you'd brought in some existing code, and a box confirming you understood you couldn't resubmit. I asked people to make their decks so they could "just finish it off within two slides."

In the order submissions came in, the six VCs each tried the demo on their own laptops, looked at the deck, and pressed PASS or FAIL. There was one criterion: do I want a follow-up meeting with this team? All a team had to do was make a VC think, the moment they saw it, "I really want to know more about this team." Votes were counted in the order they arrived, and the first side to reach 3 votes decided the result: 3 PASSes meant the final, 3 FAILs meant out. With six VCs, results usually landed within the fifth vote. Teams couldn't see which VC voted which way. Only the VCs' one-line reviews appeared on screen, without names.

It looks like majority rule, but in practice the order of the votes decided the outcome. Since it ended the moment one side hit 3, the sixth VC's vote often had nowhere to go, and even at 2 to 2, the result depended on who pressed what next. Teams that submitted earlier got judged earlier, so submission order moved results too, and in the afternoon we changed it so result announcements and finalist slots were also assigned in submission order.

The moment all 8 final slots filled, the hackathon was over for the day. Submit early and you get judged early, but submit before you're ready and get cut, and that's it. Keep polishing until 5 and there might be no slots left. That's why I warned in the opening, "It might end at 3."

It's a global hackathon, so it'd be a shame to pitch in Korean. In the opening I asked that the Shark Tank pitch, at least, be in English. "You don't need to be native speakers at all. You're going to run a global business anyway, so just imagine you're in the US right now and do it in English." We settled on a three-minute pitch in English and a seven-minute Q&A in Korean.

Homework from the VCs, and a wildcard

The VCs wrote the topics themselves, and we revealed them only around 9:40 that morning so no one could prepare ahead. On the mission slide, each click of the clicker flipped one card on the LED, and the moment all the cards were flipped, the topic declaration window opened automatically on the participant page. The six firms set nine missions in total, plus one wildcard.

BASS Ventures, with the motto "Making crazy dreams great" and investments in Travel Wallet and MarqVision, was judged by CEO Taeyang Lee. Its mission was one consumer AI used by 100 million people worldwide within a year. Four teams picked it: AZ, Heo-Kim-Hyeonggeon-Junwoo, Super Agent, and Magent Inc. AZ and Heo-Kim-Hyeonggeon-Junwoo made the final, and on the evening stage Heo-Kim-Hyeonggeon-Junwoo chose BASS Ventures, which set their mission, for a follow-up meeting.

AttentionX is a US fund that runs a fellowship for AI researchers and founders and has invested in companies like AIM Intelligence and Dentronic. Seokwon Hong came. It set two missions. Five teams took on a service that uses Korea's entertainment and gaming instincts to compete with Personal Agents in the US: CHOAM, Klaro, Chilling Soft, Team Hoyeonjigi, and Kim & Yang. Six took on a service that US RL eval and data companies would pay for: Arakyn, Kkyare, Naru, unknown, Return to the state, and TangibleIdea. Klaro made the final from the first mission and Arakyn from the second. In the evening, AttentionX set up follow-up meetings with two teams, Arakyn and Dongbyeongsangryeon, which had tackled SV Investment's mission.

SCHMIDT, which says it "invests in startups taking on big markets" and has backed FuriosaAI and Innospace, was judged by Jaewon Lee. The most teams of the day, seven, went for a product to save 10 million people whose jobs will disappear because of AI, and four went for a product with a UI other than a chat window, or no screen at all. Of the 11 teams across those two missions, only NERO made the final.

SpringCamp, with the motto "Where extraordinary founders bloom" and investments in Ohouse (Today's House) and Class101, sent partner Kyungpyo Ko. Its mission was a research agent for Korean startups that want overseas investment, and not a single team picked it. Even so, in the evening SpringCamp set up a follow-up meeting with Team Sprint, which tackled SV Investment's mission.

SV Investment, with the motto "Stand An Egg" and investments in Rebellions and HYBE, was judged by Taehong Park. Instead of a specific topic, it sent four investment criteria: an item that automates human work and has an economic and technical moat; an item with proprietary data in one vertical, like manufacturing data; an item with the same demand in other languages and cultures, so it's easy to take abroad; and an item with no clear competitor, so it can grab the market first. Those criteria, turned into one sentence, became the mission: AI that replaces human work in a vertical with proprietary data, like manufacturing. Of the three teams that picked it, Dongbyeongsangryeon, Team Sprint, and Bearing, two made the final, the highest ratio of any mission.

Goodwater Capital, a US VC that "helps exceptional founders make the world better" and invested in Toss and Coupang, was judged by Shinhye Lee. It deliberately set two broad missions so as not to limit participants' imagination. Five teams, including WJxEJ, Jae Dragon, Team Sparks, and Jongno 3-ga, took on a service for consumers worldwide seeking K-beauty and K-medical care, and Keiailab took on a global service built with stablecoins and blockchain. Neither mission produced a finalist, but on the evening stage Goodwater Capital set up a follow-up meeting with Klaro, which had tackled AttentionX's mission.

These are topics the VCs who set them are always hunting for, so picking a VC mission gives you better odds. Still, I figured some teams would surely have better ideas than these topics, so taking into account the ideas people wrote in their applications, I made one more option, a wildcard. Instead of a mission, you go with your own team's idea, but you have to pin down the target overseas market down to the country and customer. We said up front that it might come across as an idea beyond the VCs' imagination, but it could also be a disadvantage if it's a business they had no interest in to begin with. The only team that picked the wildcard was Tragatije, and that team ended up taking the last final slot.

Topic declarations were due at 10:30, but we extended three times and closed at 11:10. 36 of 41 teams declared, and once declared, a topic couldn't be changed. Each declaration was automatically posted to the participant channel, and from the afternoon the LED and the live page showed separately how many teams, and which ones, were taking on each mission.

An afternoon of green and red dots

When I designed the rules, I had Culinary Class Wars (a Korean cooking competition show) and Show Me the Money in my head. I wanted to bring that tension to a one-day hackathon. So on the left of the LED, teams stacked up in submission order, and on the right, the 8 final slots sat empty. Every time a VC made a call, a green or red dot appeared on that team's row, and to add tension we even hooked up speakers so results appeared after a drumroll on a "Made the final" or "Eliminated" screen. Three VCs judged on site and three from wherever they were. To the ones not there, I wrote in the ops channel, "If you come up to the 4th floor, the whole room buzzes every time a VC votes live."

When building started at 10, commits began piling up in every repo. The 41 teams made 2,812 commits that day. The team with the most was Klaro, with 446. Around 11:45 we handed out Cry Cheese Burger meals on the 4th floor, and everyone ate in front of their laptops while building. I expected people to submit quickly, but once it started, everyone was intensely watching to see who'd go first, and the first submission didn't come in until 2:34 p.m., four and a half hours in. It was CHOAM, which built Nunchi, an "agent that reads the room" (nunchi is the Korean art of sensing others' moods): it reads the face of someone watching a video through the camera and speaks up first. One VC wrote, "Impressive how it creatively uses the camera for the AI agent to judge the user's context," but after reaching 2 PASSes, FAIL hit 3 first.

From then on, word spread that the competition was brutal, and around 3:30, watching FAILs pile up first, I honestly worried the 8 slots might not fill. Still, I didn't want to step into the VC judging or lower the announced 3-vote threshold to 2, so I left the rules alone and wrote to the VCs in the ops channel, "Even if I end up filling the 8 teams through a second-chance round, please feel free to judge honestly." At 4:17 we also set a second-chance rule and announced it in the participant channel: if slots were still open after the deadline, eliminated teams with at least one PASS would move up in order of most PASSes, with earlier submission breaking ties. In the end all 8 slots filled, so we never needed it.

We announced results in submission order, not in the order teams passed. Even if a later team got 3 PASSes first, it waited on screen as "Results coming soon" until the earlier team's result came out. FAILs were announced as soon as they hit 3. When several results landed at once, we showed them one team at a time, 22 seconds apart, so there was time to clap.

- 2:34 p.m. First submission. CHOAM's Nunchi came in as number 1. - 3:07 to 3:38 Nine teams submitted within 30 minutes, pushing the submission count to 11. - 3:45 Twelve teams had submitted, but not one of the 8 final slots was filled. - 3:48 The day's first verdict was an elimination. The second team to submit got 3 FAILs. - 3:56 to 4:01 CHOAM got as far as 2 PASSes and stopped at 2 to 3. With the third team to submit also out, the first three results were all eliminations. - 4:05 First finalist. Arakyn took the first slot with 3 PASSes and 0 FAILs. - 4:15 NERO took the second slot, 3 to 1. - 4:36 As soon as the team before it was announced as eliminated, Dongbyeongsangryeon's result came up next, and I took the mic: "Three pass, two fail, you made the final by a hair." That was the third slot. - 4:36 to 5:01 17 teams submitted in the 25 minutes before the deadline. - 4:47 Team Sprint hit 3 PASSes first, but AZ, ahead of it in order, didn't have a result yet, so the screen showed "Results coming soon." - 4:52 AZ's votes went back and forth between PASS and FAIL to 2 to 2, and stayed frozen there for 10 minutes. - 4:53 Meanwhile, Klaro took the fourth slot, 3 to 2. - 5:01 The last submission came in, and submissions closed at 36 teams. - 5:02 AZ's last PASS came in, deciding it 3 to 2, and together with Team Sprint, which had waited 15 minutes, the fifth and sixth slots filled at once. - 5:25 Heo-Kim-Hyeonggeon-Junwoo took the seventh slot, 3 to 1. - 5:36 Tragatije, the wildcard team, took the eighth and final slot, 3 to 1. The final closed, and judging ended for the 8 teams still waiting on results.

The order of votes kept things tense to the end. Only two teams, Arakyn and Team Sprint, went straight up with 3 PASSes and no FAILs. Dongbyeongsangryeon got two PASSes, then two FAILs in a row, and survived on the fifth vote. Klaro's first vote was a FAIL, and it turned that around. AZ held on at 2 to 2 for 10 minutes before going through. On the other side, three teams including CHOAM got two PASSes but saw FAIL reach the third vote first and went out 2 to 3, and seven teams got a PASS on the first vote, got their hopes up, then took three straight FAILs. One team was frozen at 2 to 2 when the 8th slot filled.

Of the 17 teams that rushed in between 4:36 and 5:01, the 8 that submitted after 4:54 got no result because the final filled before all the VC calls came in. It was the moment the rule showed itself plainly: submit early and you're not ready and get cut, submit late and there's no room. That day, the six VCs cast 121 votes and left 114 one-line reviews. Looking back, the daytime build was far more alive than I'd expected, because the scenes of competing, failing, and passing in real time stacked up on screen all day.

An AI hackathon run by AI

It started around 11 a.m. the day before, when I pasted the event page link into Claude Code and said I needed a deck for the LED. It didn't take long to get to the point that showing passes and eliminations in real time meant PowerPoint wouldn't work and we had to build the screens ourselves. Before building screens, Claude and I worked out the rules in detail. Questions came up one by one: how to check the rule that only today's work gets judged, whether to show results the moment 3 votes come in or when I press a button, how much of the vote counts and one-line reviews to make public. Every time we settled an answer, it went straight into the rules and the screens.

One idea was to open slots at 1, 3, and 5 p.m. so that if all 8 slots filled in the morning, the other 90 people wouldn't be left with nothing to do all afternoon, but in the end we went with first come, first served. The partner logo strip on the LED and every presentation screen, since there was no good place to hang logos on site, and the kit bundled so it's easy to hand to AI, since participants would make decks with AI, also came out of these conversations.

That's how the on-site LED, the ops console on my laptop, the judging screen each VC opened on their own laptop, the submission page where participants declared topics and submitted work, the live status board, and the stage timer were all built with Claude Code the day before, and all ran off the same data all day. In the evening I went into one VC's judging screen myself and voted to test it while fixing things. The night before, we ran a rehearsal of 33 items, from finding your team, topic declaration, blocking duplicate submissions, automatic 3-vote decisions, and result announcement effects, to commit tracking, the Shark Tank timer, and matching, before the event.

At 1:57 p.m., with no submissions yet, I said the status board was too quiet. In the middle of the event, AI put up a development status screen on the LED in five minutes, showing the number of commits today, the number of teams that committed in the last 15 minutes, the commit leaderboard by team, and the latest commit messages. Two minutes later it built a screen showing how many teams were taking on each mission. At 2:11 it added a banner with a strip and alert sound at the bottom of the screen whenever a new submission came in, plus a deadline mode that locked the screen on the countdown from 4:30. At 2:30, since the text on the 16:9 LED was too small on phones, it opened a separate vertical live page, and it scheduled submission status announcements on Discord for 3:00, 3:45, 4:15, and 4:40, filled in with the real numbers at each time.

At 3:14, mid-judging, it replaced text like "VC 1/6" with dots that lit up for each VC that had decided. When someone said orange dots made it unclear whether it was a PASS or FAIL, 8 minutes later it changed PASS to green, FAIL to red, and VCs who hadn't decided to gray. From 3:16, results were announced the moment 3 votes came in, without me pressing anything in the console, and at 3:24 it changed results and finalist slots to follow submission order. At 4:25, because VCs judging in whatever order they liked meant later teams' votes were piling up first, it fixed the judging screen so VCs could only vote on the 3 earliest-submitted teams not yet decided. A request to change the word for the final round on screen took one minute to apply across the LED, the live page, the judging screen, and the automatic Discord messages.

AI was also the first to answer participants' questions on Discord. The Bloom Ops bot, originally built to collect feedback on the Bloom app, became the ops bot for the day, answering questions and posting announcements, and a separate bot confirmed receipt of submissions. When a question came in, Bloom Ops, tagged as an AI answer, replied first based on the rules and the purpose of the event, and only asked me about questions where the call was unclear. When someone asked if they could build only the deck in Lovable, it first replied that it had passed the question to the organizers, then after checking added an answer: build the deck in a way that can use your Claude credits, and use the kit. Once I settled an answer, it went straight into the bot's guidelines, and it never asked me the same question again.

For the 65 questions that came in that day, the median time for AI to post a first answer was 31 seconds, and 52 got an answer within 2 minutes. In the morning while I ran the opening and the fireside chat, and in the evening while I was on the Shark Tank stage, the bot looked after Discord, and I only had to check the summary it posted every 30 minutes. When we flipped all the mission cards on the LED, before I said anything, a post laying out each mission was already up on Discord, along with a reminder to declare topics by 10:30. The bot also posted the extension notices each time we pushed back the declaration deadline, and the 11:45 notice to pick up lunch on the 4th floor.

The rules were refined throughout the day as questions came in. At 10:12, someone asked if they could use open source, and within one minute a notice went up: allowed, but don't fork, and credit the source in the README. At 10:33, asked whether they could use another company's judgment model like JEV, the answer was that other AI models weren't allowed, and if you need judging or classification, build it with Claude's lightweight model. At 10:41, someone asked whether non-LLM AI services like STT or TTS could use external APIs. The bot first replied it would check with the organizers, then posted the conclusion four minutes later: APIs for voice and image conversion that Claude can't do directly are fine, but core AI functions like judgment and generation must be built only with Claude.

At 10:52, when someone said topic declarations were getting cut off by the character limit, a minute later the limits went up to 500 characters for the idea and 300 for the target market. Around 3 p.m., when a participant wanted to include mock data, the demo instructions field went up to 1,500 characters. An explanation for a participant's question about what happens at 3 to 3, and tips for VCs on how to look at demos, were added as announcements each time the questions came up.

The bot posted 205 messages in the participant channel. On top of 72 answers to questions, it automatically announced each topic declaration (38 times) and each finalized result (28 times). At 3:45 p.m. it posted a status update: 12 teams had submitted so far, all 8 final slots were still open, there was 1 hour 14 minutes to the deadline, and teams that were ready would be better off submitting now. Results went up on Discord, with PASS and FAIL counts, the instant they appeared on the LED with sound effects. All 28 verdicts that decided final or out were made automatically by whichever side hit 3 first, and not a single one was changed by hand by the organizers.

In the ops channel, shared by the VCs and organizers, the bot posted 83 messages. It pinged all the VCs every time a submission came in (36 times), posted at 10:38 which teams had declared for each mission, and kept up with notes on judging rules, tips for looking at demos, and reminders about judging order. From 3:16, we switched it so that the moment 3 VCs had decided, the result was announced directly on the on-site screens and Discord without organizer confirmation. During the evening Shark Tank, an alert popped up for each team saying offers were now open, and the VCs moved to the Shark Tank offer tab on their judging screens. When a VC joined the ops channel through an invite link, the bot gave them permissions within a minute or two. That was also set up in advance.

On my laptop's ops console, 11 buttons along the top switched the LED among the waiting screen, opening, topics, live board, commits, mission status, break, Shark Tank, matching, announcements, and closing. On the left, teams stacked up row by row in submission order with their status, the six VCs' votes, and the decision, and the commit warnings I mentioned sat next to team names. On the right were the remaining slots, counts of submissions, eliminations, and finalists, a preview of what was on the LED right now, and buttons for slides and Shark Tank. I'd also built a button to change a result to PASS or FAIL directly and a field to cast a vote on a VC's behalf, but there was never a need to change a verdict by hand.

Each VC opened the judging screen on their own laptop through a separate link, with their company name and logo at the top so they could tell at a glance it was their screen. The left side listed teams in submission order with status. The middle had team info, mission, target, demo instructions, demo and deck buttons, and commit history. The right had PASS and FAIL buttons, a one-line review field, and a notes field. One-line reviews were shown on the on-site screens without names, and notes were visible only to organizers. The 35 submissions with repos were locked to the commit at the moment of submission, and automatic backups ran every 5 minutes, 271 times in all.

After the event that night, using the same data, we built an archive of the 114 VC one-line reviews, grouped by result and mission without VC names. AI also drafted 11 result reports for the six VCs, Anthropic, and the Seoul city government, plus thank-you and introduction emails, which I reviewed and sent. That's how a handful of organizers could run a one-day event at this scale.

This was my first proper hackathon too, so I prepared as much as I could in advance, and talking with AI in particular surfaced things I'd never have thought of. We built a separate ops console, split the screens for the LED, VCs, participants, and organizers, added real-time features and Discord integration, and split the bots into one that takes questions and one that logs submissions. All sorts of things happen at an event, but thanks to that, I think we were able to enjoy it to the end.

Anyone can run this hackathon now

The program we used that day is open source as bloom-hackathon-kit, under the MIT license so anyone can take it and modify it. We took the program that 105 people and six VCs actually used for a day, stripped out the event settings, removed secrets like admin keys and bot tokens, participant data, partner logos, and on-site photos and videos, and set it up so you can launch it right away with demo data.

Over the day this program ran, 960 people applied and 105 came together in 41 teams. Those 41 teams' repos got 2,812 commits, and on the 36 submissions, six VCs made 121 calls and left 114 one-line reviews. The Discord bot posted 288 messages across two channels and answered 65 questions with a median first-response time of 31 seconds, and of the 28 verdicts that decided final or out, none were changed by hand. Of the 8 finalists chosen that way, 5 went on to follow-up meetings with VCs.

Participants find their team on the submission page using their application email, declare a topic, submit their work once, and watch remaining final slots and judging status on a public live page. VCs open the judging screen through their own link, look at demos, decks, and commit history, and leave PASS or FAIL and a one-line review. In the evening, they raise their hand for teams they want to meet on the offer tab of the same screen. On site there's an LED board that shows results with sound effects and a stage timer. Organizers handle scene changes, the deadline, and confirming matches from a single ops console, and after the event you're left with an archive of one-line reviews collected without VC names. The kit also includes the server API that handles submissions, voting, matching, and commit tracking, the Discord scripts that announce results automatically, and the operating approach and rules docs for having Claude answer Discord questions first.

You can also take the rules as they are: everyone competes with the same credits, code only goes in a public repo created that day, each team submits once, VCs judge in arrival order and the first side to 3 votes decides, and the door closes the moment the final slots fill. On the evening stage, VCs raise their hands offering a next meeting instead of an investment, and if several step up, the team chooses. At a time when people say hackathons have gotten boring because building got easy, I wanted to make a hackathon where you can't let your guard down all day and you walk away with real meetings.

Click "Deploy with Vercel" in the README and the repo is copied to your account, and a website, a Redis database, and storage for decks are created in one go. The only thing you need to set is the ops console password, and each VC's judging link is generated automatically from that password and shows up in the console. I hope other communities, schools, and companies can run a one-day hackathon like this with one click.

A stage that asks for the next meeting, not an investment

At 7 p.m., in Shark Tank, the 8 teams took the stage in front of the six VCs in the order they made the final. Each team presented a two-slide deck in English for three minutes, then did seven minutes of Q&A in Korean. Since they had to show both the deck and the demo within three minutes, a team member ran the demo directly on the ops laptop, I handled slide changes and screen switching from the console, and the remaining time showed on a timer on a tablet at the front of the stage. This stage wasn't about getting promised an investment. It was about booking the next meeting. After Q&A, each VC pressed "I'd like to meet" or "Pass" on their laptop, the logos of the VCs who wanted to meet appeared all at once on the LED, and if several stepped up, the team picked one. It was set up so the VCs could flip things around and pitch the team on why they should go with them.

This format, offering on the spot after each team, was decided 30 minutes before the stage. Around 6:30, while rehearsing with the console, I asked whether we'd match right after each team's pitch and move on. It turned out it was built to match all at once after all 8 teams finished. I asked to change it right there, and by 6:37 the console, LED, and VC screens had all been updated.

At first each VC could match with only one team, but during rehearsal it bothered me that a VC who matched early couldn't even make an offer to the teams after that. So at 6:45 I asked to lift the limit, and five minutes later a VC could match with multiple teams.

While we waited for one VC who was running a little late, I asked the finalists for their impressions, for the benefit of the VCs who hadn't seen the daytime energy. A member of Team Sprint said this:

Honestly, I thought you might just score things quickly, but reading the feedback one by one, every comment was thoughtful. It was a good experience where I felt that even just getting this feedback was worth it, and building with my teammates produced a much better result than doing it alone.

Arakyn said they named the team with a slight twist on Arakyd, the droid-making company from Star Wars, and that two friends who already knew each other brought a business idea they'd been talking about. They were also the first team to make the final, so I thanked them for lifting the mood when I'd been close to getting nervous early on. NERO was a team of people who'd become friends through various communities and decided to enter together.

Before the stage opened, the VCs introduced themselves in turn. After Taeyang Lee of BASS Ventures, Seokwon Hong of AttentionX said that AttentionX is a US fund focused on US companies, but any Korean founder seriously considering expanding to the US is welcome to talk anytime. Jaewon Lee of SCHMIDT introduced it as a firm that supports every round from pre-seed to IPO.

Partner Kyungpyo Ko of SpringCamp said that for the past two or three years they've had a local entity and team in the US to invest in companies expanding there, and added, "Picking eight teams was really hard." Taehong Park of SV Investment said the teams that delivered great presentations in a short time during the paper review were moving, and would treat this as a chance to learn about new ideas. Shinhye Lee of Goodwater Capital said the firm is headquartered in the US, invests mainly in consumer tech, and manages about $5 billion, cited Toss and Danggeun (Karrot) as portfolio companies, and said let's wrap this up fairly and have fun.

At 7:12, after the VC intros, the first team, Arakyn, took the stage, and the deck didn't show up on the LED. When we lifted the matching limit at 6:45, part of the code the console's deck and demo buttons and slide controls depended on had been deleted too. I asked Claude to fix it immediately, and while it looked for the cause, Claude opened Arakyn's deck PDF, which it had open in Finder, directly on screen so the pitch could go ahead. Six minutes later, at 7:18, the fixed console was redeployed. Anyone could see it was a technical problem, but I heard it went over pretty smoothly on site.

Agents that search, choose, and make the call for you

Arakyn's Agent Catalog is a search engine just for AI agents. Ask an agent to buy black headphones under $100 and it wanders through tons of websites. But websites are full of heavy data agents don't need, like CSS files, high-res images, animations, and video, so tokens leak everywhere. Agent Catalog collects product information from platforms like Taobao, Coupang, and Amazon in advance, organizes it so agents can easily understand it, and serves it through an API. When they ran it, it used nearly 30 times fewer tokens than searching with a browser. It makes money through a paid API.

Arakyn is a two-person team that picked AttentionX's mission, a service US RL eval and data companies would pay for. In the "what we built today" field they wrote that everything, from choosing the idea to development, started from zero. The demo instructions were to run a real agent searching with a web browser side by side with one searching through Agent Catalog and compare the results. The first slide split problem, solution, and business model into three columns: free for regular agent users, sold as an API subscription to companies that need bulk data. The second slide said they'd collected 300,000 product records in about three hours. By the evening pitch, that number was over 500,000.

Asked why they started with commodity goods and parts, Arakyn said that the more standardized something is, like a 30 mm nut, the harder it is to find online but the easier it is to sort as data. Questions came up about what they'd do if a big company like Google or Shopify opened similar data to agents, and about the legal issues of crawling. They answered that bundling data from many platforms is different from one company opening its own data, and that since they don't sell goods directly but hand off to purchase links, there's room to work things out with data providers. Manufacturers already have set suppliers, so the plan is to start with individuals, like someone building a drone who needs a specific part, and as agent traffic grows, get manufacturers who want more sales channels to come list their products on their own.

A team member who collects data and builds hardware personally said they'd spent a lot of time wandering foreign sites for parts that are hard to find in Korea. Product specs don't change much, so once you organize them, you can keep using them, and they saw that as a strength. In the first round, one VC had written about this team, "A data service for an endlessly growing number of agents. Very likely that agents will be buying products within the next few months." In Shark Tank, AttentionX pressed the button saying it wanted to meet.

Klaro opened its pitch with a question: "Who's your bias?" Your bias (your favorite) might be a K-pop group, an actor, or a pro gamer like Faker, and fans want to go where their bias went. But the moment they land in Korea, they run into Korean booking platforms, and AI agents that charge a monthly subscription stop at the border. Klaro's Bias is a K-fan travel agent that lets you walk the path your bias walked, like a quest.

Klaro is a three-person team that picked AttentionX's mission, a service that uses Korea's entertainment and gaming instincts to compete with US Personal Agents. They made 446 commits that day, the most of all 41 teams. In the submission form, under what they built today, they listed quest-style onboarding, a Claude agent, a VM sandbox, and Korean-language phone booking. The deck explains that a single Claude plays three roles: a personal agent that builds quest lines from public fan pilgrimage sites and Korean map data, an operator that handles Korean services on the traveler's behalf inside a traveler-only VM, and a voice agent that calls restaurants in Korean. It noted that of the roughly 18.5 million visitors to Korea in 2025, 42% said K-content made them think of traveling, and said that over the next 7 days they'd add a real phone line and a waitlist and open a beta with K-fans visiting Seoul.

On stage, from a single sign seen in someone's Instagram story, they found Yujeong Sikdang, where BTS ate as trainees. It only takes reservations by phone, so Bias called in Korean on the user's behalf. It was a demo connected to a mock restaurant, but they showed the flow: 7:00 is full, 7:30 works, one tap books it, and it carries through to Google Maps. Most travelers don't pay a monthly AI subscription, so Bias is prepaid for the days you're in Korea, like an eSIM. You need one app instead of ten local ones, and they plan to sell it where travelers already buy things, like eSIMs and rental cars. There are 225 million K-fans worldwide, and they come back to Korea for every comeback and concert.

Q&A started with how the founder managed without this service. Klaro answered that it's a travel-period-only agent for people who don't use AI often but pull out maps or Gemini only when traveling, handling calls, bookings, directions, and even talking to taxi drivers for them. It can be used from two days before departure, so you can book restaurants or exhibitions that only take phone reservations in advance. Asked whether popular restaurants have any reason to take a call from an AI, they said foreigners can't call at all, so being able to book by phone at all is already better. To a question in English about how it would fit into the existing economy where restaurants pay idols and influencers to drive visits, they said they're thinking of expanding to influencers and offering restaurants ads like search ads, with a discount for booking through the app.

In the first round, this team got both "A service that gets you past the barriers, not just a guide" and "This business model was launched as a similar service three years ago and was acquired." In Shark Tank, four VCs asked to meet at once. I flipped it around and opened time for the team to ask the VCs questions, and Klaro said the service would expand beyond K-culture to China, Africa, and other Asian cultures, so they'd like a partner who could help with that. Out came, in turn, the know-how of investing in consumer services worldwide from a US headquarters, local staff based in Boston and Singapore, and assessments of the team's model and product capabilities. A VC who has helped Korean startups expand to the US said this:

Looking only at K-fans, it's not compelling yet, and a travel agent alone isn't compelling either, but put them together and it's compelling.

Klaro chose Goodwater Capital. It was the first time that day several VCs had converged on one team.

The pitch from Tragatije, the wildcard team that took the last final slot, began with a question. What happens to a small Korean brand when a tourist who fell in love with it flies home? The answer: "It loses her forever." 18.9 million tourists visited Korea last year, and once they go home there's no way to buy again. Product pages are in Korean only, nobody answers in English or Japanese, and an agency can cost up to 1.8 million won a month.

So when a seller uploads one video, AI makes three short videos in seven languages, and if you call the number in each video, an AI agent named Tije answers. On stage, someone called in Japanese wanting to buy a knit sweater, Tije checked the product and texted a link to the product page, and the presenter added that the seller did nothing the whole time: "She was sleepy." That afternoon they'd actually asked a seller and heard there was demand, and confirmed that around 4,000 small brands in Seongsu, Myeongdong, and Hongdae are getting inquiries from foreigners. It's a subscription starting at 29,000 won a month with no commission, calls cost around 100 won a minute, and the next step is a pilot with 10 shops in Seongsu.

Tragatije is a four-person team, and their first-round result was 3 PASSes and 1 FAIL. Following the submission's demo instructions, one sample button gives you a real pre-run AI result in under 15 seconds: a script, three short videos, and copy in seven languages. Write one line about your own product or upload a vertical video, and a real AI short is made one to two minutes later. Call the number on screen and Tije actually answers, telling you about the product, price, and where to buy. The deck laid out an 805 billion won annual market based on 702,000 online sellers in Korea, 140,000 of them already using video, and a next target of 215,000 small businesses in the US. It said Claude Sonnet 5.5 handles highlights, translation, subtitles, and Tije's answers, with Tije answering only from data the seller provides.

A team member who does live commerce themselves said it bothered them that a live video, once made, disappears after one use, and they took a hint from Chinese live commerce, where offline stores filmed their own broadcasts during COVID. In Q&A, someone pointed out that many small merchants don't even have online sales channels or payment methods, so getting to an actual purchase might require something like escrow. Tragatije answered that they'd put a purchase link in the text, and for shops without one, bring them onto the platform and add payment and shipping directly. The plan is to start in Korea and expand to Japan, Southeast Asia, and the US. In the first round they got the review "As commerce channels get lighter, I expect this could become a marketing tool individual sellers can easily access," but it didn't lead to a follow-up meeting that day.

Heo-Kim-Hyeonggeon-Junwoo's Onehouse started with a question: have you ever seen friend suggestions on Facebook or Instagram? The presenter said they'd never once clicked one, because they don't know who those people are. Social media today requires you to make the first move, whether following or sending a connection request, and leans toward strengthening relationships with people you already know rather than making new ones. So they started from this: if a friend who knows you well knew everyone in the world, wouldn't they introduce you to ten people directly instead of handing you a list of ten?

Open the link and you enter a room with no sign-up or profile, and an agent called one, which knows everyone's information, speaks first. Ask it to find a designer for a side project and one scans the 290-some people in the room, brings up the best fit, and explains what that person is looking for right now and why they match you. The plan is to start with builders among US college students aged 18 to 24 looking for co-founders, teammates, or new friends, then expand to all of Gen Z.

Heo-Kim-Hyeonggeon-Junwoo is a two-person team that picked BASS Ventures' mission, a consumer AI used by 100 million people worldwide within a year, and wrote in the "what we built today" field that they built every part from scratch. The deck pointed to numbers from the FTC's lawsuit against Meta, that friends' content makes up only 17% of time spent on Facebook and 7% on Instagram, to describe social media today, and laid out how to use it in three steps. Instead of posting, you tell the agent about yourself. The agent looks at what you say and your LinkedIn and GitHub links and finds people who fit, with reasons. A conversation opens only when both sides say yes. Strangers can't message you first before you both agree.

Q&A covered whether it's a service for finding people with a clear purpose or a social network for widening your circle like LinkedIn, how it draws out intentions that are hard to put into words, and how it tracks changing interests. Onehouse said it's closer to a social network. It learns your interests and conversational style from your onboarding chat with the agent, then suggests connections over time, and it keeps information fresh by reading outside links like LinkedIn and searching the web regularly. The plan is to start with tight-knit communities, like founders looking for co-founders or designers, and grow the network as connected people link others to the people they're looking for.

In the first round this team got the review "A perspective that solves the effort of judgment through simulation." Three VCs stepped up wanting to meet, and without even asking anything, Heo-Kim-Hyeonggeon-Junwoo chose BASS Ventures, which had set their mission.

People's know-how and the field become data

NERO's Handover answers SCHMIDT's "10 million disappearing jobs" mission. Claude interviews people in jobs that AI will eliminate to draw out their know-how and turns it into sealed know-how packs, then connects them to companies and AI through blind cards with company and system names removed, leading to pack sales, per-question inquiries, and hiring. The sellers are BPO call center agents in Manila, the Philippines, who face layoffs. The industry employs 1.9 million people, and its association has lowered its 2028 employment target from 2.5 million to between 1.85 and 2.14 million. The buyers are US companies running customer service AI and AI evaluation data companies.

NERO is a three-person team. In the submission form they wrote that the whole flow was built after 10 a.m. in a new repo: the interview, Claude's follow-up questions, extraction and generalization, blind cards, demand matching, sealed Q&A, mock payment and handover viewing, and a receipt showing what Claude did. It came to 5 screens, 11 APIs, 4 Claude calls, and 10 database tables.

The first slide said "Your job may end. Your judgment doesn't." The second slide had the pricing: $149 to open a full handover, $20 per case connection, with 70% going to the author, and 15% if it leads to a hire. It also noted that the legal risk is small because only the person's judgment is uploaded, not company data, and laid out a plan to open per-case connections and line up 10 sellers and 3 buyers over the next 7 days.

In the demo, Claude added follow-up questions to the answers of a fifth-year agent at a Manila telecom and sealed the handover. For a buyer's problem, an AI that can't resolve roaming disputes from customers who changed plans mid-cycle, three cards were matched with reasons. The buyer gets answers to their questions first without opening the handover, and the full text opens only after paying. The idea split the VCs from the first round. One wrote, "A matching system centered on know-how rather than jobs felt really fresh," and another asked, "If it's a job disappearing because of AI, it's likely already been replaced by AI, so I wonder how much a know-how pack would help, and whether companies will pay."

Q&A covered how it picks the three recommended cases for each company, how it differs from existing data vendors, and in the end who's buying. NERO said that when a company enters a question, it gets categorized, AI scores the options, and the top three are recommended. As the differentiator, they pointed out that they turn agents' hands-on know-how directly into data, rather than data processed by AI. Asked whether they'd gather sellers or buyers first, they said know-how providers first, with AI scoring each provider's answers for quality and rarity. Someone also asked why this team in particular could pull off a business that requires knowing both sides well. It didn't lead to a follow-up meeting that day.

Dongbyeongsangryeon's presenter (the team name is an idiom meaning "fellow sufferers") said they had worked in shipping and logistics for 25 years and built HubTrace with an engineer with 25 years of experience, introducing the team with "So you're looking at the work of fifty years." HubTrace is a real-time 3D control tower for cargo networks. It puts cargo tracking, work orders, and delivery reassignment on one screen so dispatchers and site managers look at the same screen and act right away. In Japan, the first market they chose, truck drivers' overtime has been capped at 960 hours a year since April 2024, so 3PLs and carriers have to keep revising delivery plans while going back and forth between warehouse management and transportation management systems.

Ask the AI logistics assistant on the right, by voice or chat, where your cargo is right now, and it checks and answers. Request an urgent inspection and it creates a work order and waits for approval. Tell it to hand the remaining stops to another vehicle, and on approval the route on the map changes. Every action happens only after a person approves it. It went through 14 live evaluations and 44 offline tests, and you can try it right away with no login.

Dongbyeongsangryeon, a two-person team, laid out in the deck how they use Claude. For each request, Claude Opus 5.5 produces structured judgments of choices, true or false, and scores. The server validates those judgments and runs the lookups, and anything that changes data is held until approved with one tap. In fast mode, median response time was around 5 seconds. Over the next 7 days, they planned to add Japanese screens and voice, connect real warehouse and transportation management system data, and run a pilot with one 3PL site in Japan.

Q&A first asked why they chose Japan as the first market, where integrating with 3PL systems is hard and sales cycles are long. To the point that Japan is conservative, the presenter admitted, "Honestly, it's exactly as you say," and said Japan was just set up as a simulation model, and it could apply to all of Asia beyond Korea and Japan. Asked whether adding AI to digital twins, something people heard a lot about during the digital transformation wave, makes a change big enough for customers to feel, they said that in 25 years they always hit walls in practice whenever they started something new, and that this proposal and build-up could be a catalyst for change. They added that logistics and shipping are conservative, with high barriers and heavy data, so everyone fears change, and the plan is to start with small 3PL companies, with policy support speeding up the transition.

Asked about technical strengths, the engineer said they built it with Python crawlers and Three.js, and since the judgment model responds in 70 to 500 milliseconds, hooking it up to CCTV or sensors would let the agent make calls instantly even for safety incidents like fires. In the first round this team got "Interactive 3D is so cool. A service that's a pleasure to look at," along with homework: "Future growth speed will depend on whether the solution is modular enough to supply as SaaS or needs SI optimization for each client." Having squeaked into the final with 3 PASSes and 2 FAILs, they set up a next meeting with AttentionX.

Team Sprint's Stepcast opened with a line printed on the cover of a new IKEA manual: scan this to watch the video. Even IKEA knows paper isn't enough. A customer stuck mid-assembly calls in, and that one call costs $13.50. IKEA can spend money on videos and call centers, but small brands can't. So when a brand uploads a PDF manual, Claude drafts a 3D assembly guide, the brand reviews it, and customers just scan the QR on the box and follow along with no app.

Team Sprint, a three-person team, submitted a two-page A4 write-up instead of slides. According to it, the only input that day was the 24-page PDF manual for the IKEA KALLAX, and Claude read it and identified 9 types of parts and 14 assembly steps. The panels are rendered in 3D with an oak texture, the screws and dowels are modeled in their actual shapes, each part slides into place step by step, and the hex key and hammer actually turn. They confirmed the same approach works on a much more complex model airplane kit manual. On the review screen, changing the screw in step 3 from part 1 to part 2 updates the parts list and quantities for the remaining steps. AI makes the draft, and the brand that knows the product best is responsible for accuracy.

People already pay for this. Bilt, which makes 3D guides by hand, takes 3 to 6 days per product, and one of its clients saw support calls drop 30%. Wayfair alone has 20,000 suppliers, most of them small companies. AI runs only once per product, so there's no AI cost when customers view it, and the new Claude model cut costs by 40%. The moat is data on which step customers get stuck on, and when a brand fixes that step, the QR stays the same, so nothing needs reprinting. The pitch ended with "Paper is not enough, not even for IKEA."

Asked why companies wouldn't just make 3D with AI tools themselves, Team Sprint admitted Claude models are open to anyone and anyone can do 3D modeling, but said they revised the first model more than 20 times that day and reviewed it by eye like an actual product manager would, so the quality itself is different. Asked about go-to-market, they said a prototype takes two to three weeks, so they'd start by meeting CEOs of small and mid-sized companies that are required by regulation to include manuals, and if things went well here, they'd like to be connected to IKEA too, finding their way step by step. In the first round this team got no FAILs and the review "A fresh item with possible future M&A potential. However, growth speed will be determined by the technical barrier to entry and the AI token cost structure for 2D to 3D conversion." In Shark Tank, SpringCamp asked for a next meeting.

AZ's Story Studio took BASS Ventures' "100 million users in a year" mission into web novels, where the 100 million are readers, not writers. Throw in notes, photos, long writing, or documents, and AI interviews you to grow a story knowledge graph. Agents playing planner, writer, editor, reader, and cover designer use that graph as memory to co-write web novels and drama scripts that don't fall apart even as they get long. Choose a direction at a fork in the story and the editor catches continuity errors while the reader agent predicts reactions like "cider" and "sweet potato" (Korean slang for satisfying and frustrating plot turns). Hit publish and Claude draws a cover and makes it free for anyone to read, like a blog. The target is people in the US who enjoy web novels and short dramas but have never written. In the first round, one VC wrote this about the team:

Impressive perspective: it's not that people can't write, it's that they can't get it out.

AZ, a two-person team, set its first market in the deck as US and English-language web novel writers serializing on Wattpad, Webnovel, Royal Road, and Tapas, and its next markets as Japan's Narou and Kakuyomu, then Southeast Asia. They named the problem as continuity collapse, where characters, rules, and foreshadowing fall apart around a few hundred chapters when you post every day, and noted that every agent runs only on Claude Opus, Sonnet, and Haiku 5.5, with Korean and English screens and mobile support.

Q&A started with whether they meant to gather 100 million writers. AZ said the 100 million are readers and they want to be a platform. To the point that the interface lets anyone start easily but the content seems aimed at professional writers, they said the goal is a platform where anyone can publish content with AI's help and build their own brand as a writer. Asked how they'd establish themselves when AI-made content is flooding in, freshness is dropping, and web novel communities are wary, they said they'd lead with writing support, with an editing manager deciding direction together with the writer at each fork in the story.

Asked who owns the copyright if a hit comes out, they said the writer owns it, the platform stays in a supporting role collecting subscription fees, with ads on free works and revenue sharing on paid ones. Asked why they could do better than a similar service already in Taiwan, or Claude itself, they said they're considering lowering fees early to attract writers and giving writers all paid revenue without settlement deductions.

The two, who've written code for a long time, believe that now that the value of code no longer belongs only to developers, the next place for what AI can do is content creation. In the first round AZ also got the critique "Many similar services are already out, and the key factor is how well-structured the works it produces are, beyond simply producing them." It didn't lead to a follow-up meeting that day.

What we heard most: the rules were unusual

After all 8 pitches, we put the final results on the LED. Arakyn and Dongbyeongsangryeon with AttentionX, Klaro with Goodwater Capital, Team Sprint with SpringCamp, and Heo-Kim-Hyeonggeon-Junwoo with BASS Ventures set up follow-up meetings, so 5 of the 8 teams made next appointments with VCs. AttentionX connecting with two teams wouldn't have happened if we hadn't lifted the matching limit right before the stage. Regardless of matching, all 8 finalists got $5,000 in Claude credits per team.

Looking at the results, the missions VCs set and the teams VCs chose in the evening didn't always line up. The only pairs where a VC connected with a team that solved its own mission were BASS Ventures with Heo-Kim-Hyeonggeon-Junwoo and AttentionX with Arakyn. AttentionX also picked Dongbyeongsangryeon, which solved SV Investment's mission. SpringCamp picked Team Sprint from the same mission, and Goodwater Capital picked Klaro, which solved AttentionX's mission. The missions were only a starting point. In the evening, the reasons VCs raised their hands came down to the team and the product.

We thanked the six VCs who stayed from the weekend afternoon into the night, Anthropic for its generous credits, and Seoul AI Hub for opening its space. At the end, the VCs, the founding teams, and the participants who stayed to the end gathered together, shouted "Ambition for Korea, fighting!" (a Korean cheer), and took a group photo. That night, messages like "Great work, everyone~! The hackathon was fun~!" and "Great job, everyone!!" kept coming into the participant channel until late.

What we heard most after the event was that the program and rules were unusual. Every team's commits piled up on screen all day, green and red dots appeared every time a VC made a call, and the door shut the moment 8 slots filled, so that makes sense. The 114 VC one-line reviews, along with each team's demo, deck, and GitHub links, are collected in the archive, which you can also browse by mission, and we're connecting the five matched teams directly with the VCs.

The embarrassment of saying "going global" didn't vanish in a day. But now there are 105 people who, instead of just saying it, actually did it for a day, and five teams are now continuing the conversation with VCs. I'm not going to wait anymore.

Every service built that day

Here are all the services from the 36 teams that submitted. The 8 finalists come first, and the rest are in submission order. The one-line descriptions are the teams' own words, translated. Demo, GitHub, and deck links for every team are in the archive. Demos and repos are run by the teams, so some may go offline later.

- Arakyn · Agent Catalog (finalist, follow-up meeting with AttentionX). Faster. 30× lighter. The search engine built for agents. - NERO · Handover (finalist). A market where Claude turns the know-how of jobs AI will eliminate into sealed know-how packs and matches them with companies and AI through blind cards, leading to pack sales, per-question inquiries, and hiring. - Dongbyeongsangryeon · HubTrace (finalist, follow-up meeting with AttentionX). A live 3D control tower for cargo networks, with an AI logistics assistant that looks up, plans and acts, only with human approval. - Klaro · Bias (finalist, follow-up meeting with Goodwater Capital). A companion travel agent for K-fans that lets you walk the path your bias walked, like a quest. - AZ · Story Studio (finalist). An AI writing studio where you throw in material and an AI writers' room grows a story knowledge graph, co-serializing web novels and dramas that don't fall apart even at full length. - Team Sprint · Stepcast (finalist, follow-up meeting with SpringCamp). AI that turns a brand's PDF manual into a 3D assembly guide customers follow via QR, with no app. - Heo-Kim-Hyeonggeon-Junwoo · Onehouse (finalist, follow-up meeting with BASS Ventures). A new agentic social network that proactively connects you with strangers who fit you. - Tragatije · Tragatije (finalist). Short-form video becomes a live stream that never ends, 24 hours a day. - CHOAM · Nunchi. The agent that reads the room, and moves you (futuristic, somewhat optimistic). - Planetes · Quest. The anonymous B2B marketplace for engineering work: hide who is asking, prove what was delivered. - Keiailab · settlivo. A global payroll platform that simplifies the complicated process of paying remote workers at overseas companies and uses blockchain to cut transfer time and fees. - WJxEJ · Glowprint. Before you pay for a K-beauty product or a Seoul clinic, check at a glance whether the seller is trustworthy and whether it suits your ingredients. - Kkyare · Weather Decision Gym. A service that forecasts the weather and this year's expected yield for every field on large US farms, retraining on actual harvest records each season to get more accurate over time. - Jae Dragon · muselook. Pick any makeup look you love, see it simulated on your own face, and shop the K-beauty products that match its colors, re-picked for your skin tone. - Nam-ui Jiwon · gigi. gigi uses AI to connect businesses with on-demand talent, and workers with continuous income and portable benefits. - Magent Inc. · Magent. A hyper-personalized consumer AI that remembers your choices and optimizes your next purchase. - unknown · Into the Unknown. A de-identification service for mental health data that preserves clinical meaning and data quality. - Triple Hyun · WGS (worktree guild system). "Guild," a job-sharing community for 10 million people working in the US tech industry. - Bearing · Bearing. AI that writes the customs paperwork for everything that goes onto a ship. - Chilling Soft · Alter. A personal agent that lives at the OS layer, not in a chat window. It quietly learns you, helps before you ask, and becomes more like you the more you use it. - Team Kimbakgui · Dream Achieve Moments (DAMs). AI can replace your tasks, not your expertise. Dream Achieve Moments reads the life you've already lived, finds where your skills and your dream meet, and gets (the description cuts off here in the original). - UJHS · SCENE. A 3D messenger that turns your group chat into a round table. Friends who are reading sit up, friends who aren't are slumped over, and your words are delivered in your own voice. - Naru · Frontier Lab Market. A market that turns field experts' tacit knowledge into rights-verified AI training and evaluation data and virtual companies for LLMs. - Team Hoyeonjigi · Lumi. An AI agent that gathers news about your favorite K-pop artist in one place and connects you to their outfits, merch, and favorite items so you can buy them in the US too. - SEVENFOLDER · PathPilot. AI legacy patch intelligence. Even systems that can't be upgraded can respond on day one. - Return to the state · RLMask. A persona-based de-identification service that runs locally, with no risk of personal data leaking. - Kim & Yang · YuiYumi. A Live2D character AI assistant that lives on your screen, watches what you do, talks to you first, and handles everything from email and scheduling to building apps for you. - Newtype · GhostReel. A global short-form side-hustle channel where a twin creator films faceless videos for you based on viral short-form data and your own sense, and your Claude reads and rewrites through MCP to make them better and better. - Team Choigo Baekgu · Morrow: There's More to Tomorrow. A service where your AI agent interviews with companies' hiring agents first and finds roles your resume doesn't show. - Grab Credit Tao Jayul · poin. Chat answers what you say. poin answers what you don't. - Team Sparks · Jinseonmi. An AI guide that lets foreigners and tourists see local Korean information with sources, get K-beauty and K-medical comparisons and personalized recommendations, and build a visit itinerary. - Jongno 3-ga · Pibook. An AI-agent-centered platform that picks and manages the best procedure, price, and clinic purely through autonomous agent-to-agent (A2A) negotiation and automatic matching among foreigners, Koreans, and clinics. Sensitive procedure data and clinic quotes are protected with zero-knowledge proofs (ZK) (the description cuts off here in the original). - Team BuleMoonDream · SIEVE_WHYBUY. Now that the leading frontier companies have all entered the agent market, SIEVE_WHYBUY connects what consumers actually bought after an AI recommended a product, and why they chose something else or gave up, giving consumers rewards for participating and giving brands post-recommendation purchase decision data. - Super Agent · Meme Rush. Every day, agents turn the world's hottest meme and hit game into a 30-second game and deliver it through social media. - TangibleIdea · Argos. Vendor access to your data: attack-proven, set up by AI. - animalive · Companion Studio. A Live2D companion you talk to, not type to. She answers in her own voice and reacts with her face, every turn.

Originally written in Korean on Brunch, October 25, 2026.