c0mpiled_
← back to articles
hack·july 13, 2026·30 min read

gstack x gbrain Hackathon #2 (San FranOsaka)

by james tan445 reads

Overview

  • 238 registrations, 200+ participants, 60+ submissions (55 unique projects), 11 prizes awarded.
  • Hosted at Ritsumeikan University in collaboration with Osaka University.
  • Landed on the tails of IVS in Kyoto.
  • Many first-time hackathon attendees, including a strong student contingent.
  • Many attendees from abroad — Malaysia, Singapore, Taiwan, and beyond.

Reflections

Good

  • The surprise attendance of Garry Tan and Sean Grove was great and energized the room.
  • Keeping the building section physically close to the final presentation section proved good for logistics.
  • We nailed timing pretty well across the board.
  • Incredibly strong submission quality — we had been battling disqualifications pretty heavily in the past, but this was the best turnout of high-quality projects so far.
  • Teams were exceptionally articulate in describing their projects.
  • Talent density in the room was very high — many participants were already running complex, sophisticated Claude Code setups and harnesses.
  • Everyone in the room was really familiar with YC and startups, and many seemed eager to come to the US.
  • Met quite a number of first-time hackathon participants and students.

Bad

  • Late food, and not enough of it. We ordered for approximately 70% of the field; ordering for 100–120% of the field is best.
  • The stage collapsed!

Neutral

  • We experimented with the idea of having ecosystem partners, which was new — we generally don't partner with others unless the alignment is super strong.

Sponsors & Judges

200+ participants, 60+ submissions, 11 prizes awarded! Amazing participant-to-submission ratio, some of the strongest teams and projects we've seen to date, the best technical talent — all supported by the venerable Sean Grove, Shoya Matsumori, Jia Shen, Jian Shen, Sina Matian, Yad Smood, Seiji Sakurai, and of course Garry Tan.

Prizes & Recognition

1st Place: KashinAI — KashinAI is a macOS assistant that rapidly parses complex context from your current screen — company knowledge, customer state, history, local memory — and turns it into a single sentence that moves the situation forward "right now." It fuses on-screen text, accessibility data, browser information, and screenshots/OCR into a self-improving learning loop, ultimately functioning as a one-key "clear this writing task" button.

2nd Place: SenpAI — SenpAI is a company brain for manufacturing. It compiles the design knowledge buried in drawing revision histories (△ marks) and trouble reports into executable checks, and interrupts designers at the moment they're about to repeat a past failure — like a senior engineer watching over your shoulder. Every accepted fix feeds a new structured lesson back into the brain, so the more you design, the smarter it gets.

3rd Place: Expert OS — App store and talent market for agents. Your agent directly accesses the web3 skill market and automatically charges the skill rent fee, and sends emails to registered freelancers when skills are not enough.

Best Design: Bonsai.ai — It's the Dynamic Software Interfaces thesis made concrete: say a goal in plain language, and an agent composes a living, native screen for it from a frozen, whitelisted component pool — structured DSL only, rendered on-device. Every goal lives inside PARA (Projects, Areas, Resources, Archive), so your life's context is organized by purpose, not by which tool holds it. The interface isn't generated once: as a goal accumulates context, its screen evolves with it — a job hunt starts as a day-one checklist collecting missing context, and by day ninety has reorganized into a decision dashboard.

Best Engineering: RoboTIDE — RoboTIDE turns scattered robot documentation — CAD files, firmware repos, manuals, and sensor logs — into a live knowledge graph, then fans out dozens of parallel diagnostic checks the instant something fails, instead of an engineer manually tracing a fault through hardware and software one subsystem at a time. It correlates the results, either fixes the issue autonomously or hands the engineer the exact next step, and gets smarter with every incident it logs. What used to take hours of part-by-part investigation now takes minutes.

Most Investable: Xpert Finder AI — Expert Finder AI is a middle man that helps employees find the right Person in Charge (PIC) in seconds instead of hours. Companies, therefore, can reduce time spent searching, minimize duplicated work, preserve organizational knowledge and improve collaboration — all without exposing confidential information.

Linzumi Award (1-year Claude Codex subscription for the team): GHeart — Coding agents have collapsed the cost of writing code to near zero, but reviewing still costs human time — gstack, for example, has 462 PRs open right now, including last hackathon's winning project, unreviewed for 7 weeks. GHeart cuts the cognitive load of that queue: each PR is distilled to one card — CI status, a mergeability score, an auto-generated preview, and a plain-English TL;DR — and cleared with a single swipe that submits a real GitHub review. And every swipe trains a per-reviewer brain (using gbrain!) — a dynamic software interface that keeps reshaping itself around your judgment.

GStack x GBrain Award ($500): Omnibrain — OmniBrain is the enterprise gbrain. It federates the knowledge scattered across SharePoint, Notion, Google Drive, Slack and ServiceNow. It continuously crawls your enterprise data where it lives, keeping it up to date and extracting skills. Every employee gets an email with a scoped MCP endpoint or CLI for their agents (Claude, Codex, Cursor) to plug into, so agents can retrieve up-to-date facts and the skills scoped to the work they are trying to do. Glean sells access without process; Celonis sells process without access; OmniBrain provides the intersection of both.

Jinba Award ($250): Jarvis — Agents finish in seconds, then make you read an essay. Jarvis turns any model into one that answers in UI — and lets any agent render into that UI.

Shisa AI Award ($250 + tokens): gevent — gevent is a set of AI coding-agent skills for running tech events. Its skill turns rough talk notes into a live presentation website where the speaker drives the slides and every attendee gets their own accessible, personalized view in real time — adapting language (EN/JA), reading level, font size, and contrast per person, all served from a lightweight Bun/TypeScript runtime over a shareable URL. It also ships skills to auto-create Luma and connpass event pages and to research any event into structured notes.

Outerport Award ($250): Foundry Brain — An AI yield engineer for semiconductor fabs — Company Brain, built for the fab. When yield suddenly drops, Foundry Brain walks the fab's three disconnected data systems, names the root-cause machine with evidence, recommends hold-or-ship, and replays its entire investigation in a live UI.

Submissions by Track

  • Company Brain — 33 (60.0%)
  • GStack x GBrain — 23 (41.8%)
  • Software for Agents — 21 (38.2%)
  • Dynamic Software Interfaces — 20 (36.4%)
  • CrustData — 3 (5.5%)
  • Linzumi — 1 (1.8%)

Total unique submissions: 55. Many teams tagged multiple tracks, so percentages don't sum to 100%.

All Submissions

AccessGraph (Company Brain · Software for Agents · GStack x GBrain)

AGENT-M² (Software for Agents · GStack x GBrain)

  • Description: Agent-M² is a deterministic replay and regression testing tool for AI agents. It records an agent's production run as a causal DAG of LLM calls, tool calls, responses, dependencies, and hashes, then replays the exact run locally with zero network calls or duplicate side effects. When agent behavior changes, Agent-M² identifies the exact divergent event, expected vs. actual payload, and upstream cause so teams can debug and test agents safely before deployment.
  • Team: Marvel Harisson, Moaaz Negm
  • GitHub: https://github.com/INo-xious/y-combinator-hackathon
  • Video: https://canva.link/t80wm68ug991n8i

AtoA (Company Brain · Dynamic Software Interfaces · Software for Agents · GStack x GBrain)

bad happens (Company Brain)

Bonsai.ai (Dynamic Software Interfaces · GStack x GBrain) — Best Design

  • Description: It's the Dynamic Software Interfaces thesis made concrete: say a goal in plain language, and an agent composes a living, native screen for it from a frozen, whitelisted component pool — structured DSL only, rendered on-device. Every goal lives inside PARA (Projects, Areas, Resources, Archive), so your life's context is organized by purpose, not by which tool holds it. The interface isn't generated once: as a goal accumulates context, its screen evolves with it — a job hunt starts as a day-one checklist collecting missing context, and by day ninety has reorganized into a decision dashboard.
  • Team: Guoliang Chen, Cherry Lu
  • GitHub: https://github.com/Freezind/Bonsai.ai
  • Video: https://drive.google.com/file/d/1rlCCQqM_kr8PgX9MwC__LbThJHandinD/view?usp=sharing

Brain contener (Company Brain)

  • Description: A SaaS platform that continuously observes how work is actually done through SaaS activity and browser interactions, automatically discovering and documenting recurring workflows. It analyzes workflow frequency and tools used to recommend which tasks should be delegated to AI agents, turned into reusable skills, or fully automated. This enables companies and AI implementation teams to identify AI opportunities without lengthy interviews or consulting engagements.
  • Team: Chiharu Okabayashi, Yuki Nishimura
  • GitHub: https://github.com/niyu521/worklog-analyzer
  • Video: https://company-brain-2.vercel.app/

BrainBox (Company Brain)

Callstack (Software for Agents · GStack x GBrain)

Clay (Dynamic Software Interfaces · GStack x GBrain)

  • Description: Clay is a delivery architecture — not an app — that lets a user's own coding agent compile a structurally different interface from the same developer-shipped data, live, per user and per intent. Developers publish two frozen contracts (a semantic data schema and a closed component registry), and every AI-generated interface is validated against them before it ever renders, so nothing invalid or hallucinated reaches the screen. The demo proves this on a university LMS payload — the same JSON data compiles into a minimalist student portal or a dense professor grading grid on command — but the same engine works unmodified for any structured-data product: email, CRM, or beyond.
  • Team: Taiga Takahashi, Lee JiHwan, Maksym Sydorov, Wai Yan Zaw Win
  • GitHub: https://github.com/wai-yan-zaw-win/Clay
  • Video: https://youtu.be/e4EZmBBS1d4

ContentOps AI (Company Brain)

  • Description: ContentOps AI is an AI content operations command center for marketing and sales teams. You start with a campaign goal, and it orchestrates specialized agents through research, strategy, post and video creation, approval, scheduling, publishing, lead discovery, outreach, and ROI analytics — all from one dashboard. It's built as a visual workflow product (not a chatbot), with cost-aware model routing and platform publishing adapters for LinkedIn, Instagram, X, and more.
  • Team: Farhan Morshed, Irfan Faisal
  • GitHub: https://github.com/Morshedwork/contenthack
  • Video: https://youtu.be/h54zGzuIG_g

DockMind (Company Brain · Software for Agents · GStack x GBrain · CrustData)

Docutor (Company Brain · Software for Agents)

Embedded UI QA Testing (Dynamic Software Interfaces · Software for Agents · CrustData)

  • Description: Embedded device interfaces are still fixed — if a user wants a different workflow, layout, or personality on a small hardware screen, the loop is slow: describe the change, edit firmware/UI code, build, flash, inspect the physical display, and repeat. With PlaiPin, you type one command — "pomodoro timer" — and the agent generates the interface, deploys it to the real device, captures what the physical screen actually shows, critiques it the way a designer would (legibility at arm's length, contrast, hierarchy), and deploys the improvement. Hardware-in-the-loop testing, applied to interfaces.
  • Team: Natalie Yeo, Irvin Cardenas
  • GitHub: https://github.com/PlaiPin/embedded-ui-review
  • Video: https://drive.google.com/drive/folders/1-1fPzQDMyqq0tjZfd09DVY4J7Zqw-NSe

Enroll (Software for Agents)

Eula (Dynamic Software Interfaces)

  • Description: Modern dating has turned into a social hierarchy game. People swipe for weeks based on status, looks, and money instead of actual compatibility, and the relationships that are supposed to support us end up stressing us out instead. We built a platform where you interact with a dynamic, expressive digital partner (not just a chatbot) in a shared environment you can shape and create together. That interaction reveals your real vibe and substance, which we use to match you with someone you'd actually click with, not just someone who checks the right boxes.
  • Team: Yusei Otsuka, Shimpei Kawakami
  • GitHub: https://github.com/yuseict25-debug/Hackathon-yc-.git
  • Video: https://youtu.be/I1zsI4UPUHQ

exe (Company Brain)

  • Description: exe records a meeting, transcribes it, and automatically creates tasks and their dependencies, then shows them on a Gantt chart. When a person has too many open tasks, it calls their iPhone and an AI voice agent discusses the work with them and reassigns some of it to people with capacity, along with a generated handoff note. The voice agent doesn't stop to run tools during the call — it sends the transcript to a separate agent that does that work in the background.
  • Team: Takuto Kinjo, Kentaro Ishikawa, Shino Ito, Tomohiro Takeda
  • GitHub: https://github.com/sanbasan/exe
  • Video: https://canva.link/uphc5e3ou1qlodq

Expert OS (Software for Agents) — 3rd Place

Facet (Dynamic Software Interfaces · GStack x GBrain)

Field Brain (Company Brain)

Flowork (Company Brain)

Foundry Brain (Company Brain) — Outerport Award

G Link (Company Brain · Dynamic Software Interfaces)

gbrainOpt (Company Brain · GStack x GBrain)

gevent (Dynamic Software Interfaces · GStack x GBrain) — Shisa AI Award

  • Description: gevent is a set of AI coding-agent skills for running tech events. Its skill turns rough talk notes into a live presentation website where the speaker drives the slides and every attendee gets their own accessible, personalized view in real time — adapting language (EN/JA), reading level, font size, and contrast per person, all served from a lightweight Bun/TypeScript runtime over a shareable URL. It also ships skills to auto-create Luma and connpass event pages and to research any event into structured notes.
  • Team: Sota Kobayashi, Haruto Watanabe, Sota Suyama, Haruka Takahira
  • GitHub: https://github.com/sota009/gevent
  • Video: https://gdgs.jp/c0mpile

Ggent (Company Brain)

  • Description: Ggent mines the judgment your ace employee can't explain — from the diffs between junior drafts and her revisions, joined with win/loss records — and stores it as a knowledge graph connected the way real work connects, not by semantic similarity. A new brief pulls up everything that actually mattered, and the agent outputs proposals with the ace's sense, validated against her held-out revisions.
  • Team: Ken Nagata
  • GitHub: https://github.com/yukik8/ggent
  • Video: https://youtu.be/v0Kckpwb4nc

GHeart (Company Brain · Dynamic Software Interfaces · Linzumi · GStack x GBrain) — Linzumi Award

  • Description: Coding agents have collapsed the cost of writing code to near zero, but reviewing still costs human time — gstack, for example, has 462 PRs open right now, including last hackathon's winning project, unreviewed for 7 weeks. GHeart cuts the cognitive load of that queue: each PR is distilled to one card — CI status, a mergeability score, an auto-generated preview, and a plain-English TL;DR — and cleared with a single swipe that submits a real GitHub review. And every swipe trains a per-reviewer brain (using gbrain!) — a dynamic software interface that keeps reshaping itself around your judgment.
  • Team: Yoshihiro Sekiya, Arata Jingu, Giacomo Carlo Guiulfo Knell
  • GitHub: https://github.com/giacope/gheart
  • Video: https://gheart.fly.dev/

Hackathonday (Company Brain · Dynamic Software Interfaces · Software for Agents)

  • Description: Every company runs on forms nobody designed for you. Twelve fields, eleven of which the company already knows. That's because every ERP form is a guess, frozen at design time. Policy changes — forms don't. So we stopped building forms.
  • Team: Hackathonday
  • GitHub: https://github.com/hackathonday/0
  • Video: https://github.com/hackathonday/0

Jarvis (Dynamic Software Interfaces · Software for Agents · GStack x GBrain) — Jinba Award

KashinAI (Company Brain · GStack x GBrain) — 1st Place

  • Description: KashinAI is a macOS assistant that rapidly parses complex context from your current screen — company knowledge, customer state, history, local memory — and turns it into a single sentence that moves the situation forward "right now." It fuses on-screen text, accessibility data, browser information, and screenshots/OCR into a self-improving learning loop, ultimately functioning as a one-key "clear this writing task" button.
  • Team: Toru Tano, Woojin Song
  • GitHub: https://github.com/Torutesu/KashinAI
  • Video: https://www.loom.com/share/2219b209c4364246821a6e4591ccb49f

MemoryPath (Company Brain)

MOKUROKU (Company Brain · Software for Agents)

nolan (Company Brain · Dynamic Software Interfaces)

Noren (Dynamic Software Interfaces · Software for Agents)

note OS (Company Brain)

Omnibrain (Company Brain · GStack x GBrain) — GStack x GBrain Award

  • Description: OmniBrain is the enterprise gbrain. It federates the knowledge scattered across SharePoint, Notion, Google Drive, Slack and ServiceNow. It continuously crawls your enterprise data where it lives, keeping it up to date and extracting skills. Every employee gets an email with a scoped MCP endpoint or CLI for their agents (Claude, Codex, Cursor) to plug into, so agents can retrieve up-to-date facts and the skills scoped to the work they are trying to do. Glean sells access without process; Celonis sells process without access; OmniBrain provides the intersection of both.
  • Team: Eduardo Gonzalez, Rahul Soans, Sitora Kakhramonova
  • GitHub: https://github.com/xpressai/omnibrain
  • Video: https://youtu.be/Wqdg6BOKU2c

Perfectus (Company Brain · Software for Agents · GStack x GBrain)

Pillow (Dynamic Software Interfaces · GStack x GBrain)

  • Description: Pillow is a realtime collaborative canvas where an AI teammate — with its own visible cursor — proactively facilitates technical meetings, turning live conversation into charts, tables, diagrams, and design docs as you speak. Grounded in your team's context through Slack, meeting transcripts, and our memory layer (gbrain), it doesn't just take notes — it thinks with the team, refining architecture and decisions in real time. Until now, meetings died as unreadable text minutes; Pillow removes the hardest step in teamwork — the gap between an idea spoken aloud and something concrete you can build on.
  • Team: Yusei Moriwaki, Yuichi Matsuoka, Yusuke Ikoma, Tomoki Taguchi
  • GitHub: https://github.com/cogni-engine/yc-hackathon
  • Video: https://www.loom.com/share/00a239d02cc74000821b35e4e488a8d0

ringi (Software for Agents)

RoboTIDE (Company Brain) — Best Engineering

RoundLabs (Dynamic Software Interfaces · GStack x GBrain)

SenpAI (Company Brain) — 2nd Place

Serendipity (Company Brain)

  • Description: AI agents discover cross-industry partners inside a TEE, without exposing company secrets. Random networking is broken. Serendipity lets company AI agents enter a secure match room, reason over private company context, and output only safe collaboration opportunities.
  • Team: Kyosuke Yanagisawa, Jun Kawai, AYJYANA
  • GitHub: https://github.com/kyoky0/7-5-yc
  • Video: https://youtu.be/y1dvy2W-oZ4

sigilog (Company Brain · Software for Agents)

  • Description: Sigilog is a protocol layer designed for multi-agent AI systems to establish verified identity, tamper-evident messaging, and replayable audit trails. By utilizing cryptographic signatures and an append-only hash-chained ledger, it acts as a trust and provenance layer — often described as "TLS + git for AI agents" — to secure inter-agent communications. Ultimately, this enables organizations to definitively prove exactly how autonomous AI decisions were reached and guarantee that no intermediate data was secretly altered in transit.
  • Team: Tsuda Rin, Haekal Damar Qudsi Madani, MD Tamjidur Rahman
  • GitHub: https://github.com/MelancholicMelon/sigilog
  • Video: https://youtu.be/_whZbOb-mU4

SkillBrain (Company Brain)

SkillForge (Company Brain · Dynamic Software Interfaces · GStack x GBrain)

SprintDesk (Dynamic Software Interfaces)

Superconnect (Company Brain · Software for Agents)

TBrain (GStack x GBrain)

TECTEME / Programmable Matter (Dynamic Software Interfaces · Software for Agents)

  • Description: TECTEME is building a modular design-and-build system that makes physical spaces programmable using small, reusable Octet Modules. By combining structural modules, design software, and robotic assembly, TECTEME enables spaces to be designed, generated, modified, and reused like software. The initial market is branded pop-ups and exhibition spaces, with a long-term vision to transform architecture into an adaptive system, almost like a living organism.
  • Team: Kaito Yaegashi, Kazuki Koike, Amon Suzuki
  • GitHub: https://github.com/ProgrammableArchitecturePJ/modularizer
  • Video: https://youtu.be/xeCS6Ys7X1Y

THE GROUND TRUTH (Company Brain · Software for Agents · GStack x GBrain)

TRIAL GATE (Software for Agents)

TripReel (Dynamic Software Interfaces · GStack x GBrain)

  • Description: TripReel AI is the travel video editor that understands your trip: upload raw clips, choose your story style, and it instantly turns them into a polished short reel with subtitles, places, time, budget, and route overlays. Instead of one fixed editor, TripReel generates a custom editing interface from a "video recipe," so a day trip, budget challenge, or route story each gets the right controls and visual layout. It helps travelers post beautiful, information-rich reels in minutes instead of spending hours across CapCut, maps, subtitle tools, and manual edits.
  • Team: Yasuhisa Suzuta, Tomohito Suzuta, Shutaro Onishi, Whiplus Wang
  • GitHub: https://github.com/whiplus/suzuta_brothers
  • Video: https://suzuta-brothers.vercel.app/

Vibe Planner (Dynamic Software Interfaces)

Vineyard (Earning as a Service) (Software for Agents · GStack x GBrain · CrustData)

Xpert Finder AI (Company Brain · GStack x GBrain) — Most Investable

* * *