hack·october 21, 2026·13 min read

c0mpiled-14: UC Berkeley

by Jacob Yeum780 reads

Overview

  • 184 registrations, 25 projects submitted by 70 builders, and 3 prizes awarded.
  • Teams built against the YC Requests for Startups (Fall 2026) list. Multiplayer AI was the most popular RFS with 6 projects, followed by Self-Maintaining APIs (4).
  • None of this would have happened without the support of Enactus and our partners Superset, Freestyle, Latted, Gutgutgoose, Quivly AI, Sona8, & RunAnywhere, who helped make the hackathon run as well as it did!

Reflections

Good

  • Strong founder intent: 11 of 25 teams said they are applying to YC now, and another 12 said they plan to apply later.
  • Lots of full teams: 9 projects came from teams of four, and 5 were built solo.
  • The overall quality of submissions was high. Judges were impressed by the overall output!

Bad

  • Participation fell a little shorter than we expected, but a smaller crowd made the event more intimate and collaborative.

Sponsors & Judges

25 submissions, 3 prizes awarded! Hosted by c0mpiled (by Transpose Platform) and Enactus.

Sponsors: Transpose Platform, Superset (YC P26), Freestyle (YC S24), Latted (YC W24), Gutgutgoose (YC S26), Quivly AI (YC F26), Sona8 (YC F26), RunAnywhere (YC W26), Enactus

Judges

Prizes & Recognition

Transpose Platform — YC Requests for Startups (Fall 2026)

Prizes: 1st $1,500 + $10,000 in Freestyle credits · 2nd $900 · 3rd $600

1st Place: TerraSwarm ↓ entry — TerraSwarm is a swarm of low-cost ground robots that collect dense, location-tagged environmental data where wildfires actually start, and stream it in real time to a map that pinpoints emerging hotspots. By using many cheap robots in place of a few expensive drones and satellites, TerraSwarm gives utilities, insurers, and land managers the continuous ground-level data they need to model wildfire risk and catch fires earlier.

Ethan Sie, Ryan Li, Tony Ly

GitHub: N/A | Video: https://youtube.com/watch?v=vjnJ-qpns70&feature=youtu.be

2nd Place: Penny ↓ entry — AI agents like Meta’s Muse and Instinct are already buying for millions of people, but most businesses have no way to sell to them. Penny lets any business get paid by AI agents over instant stablecoin rails, so customers get 3% off and businesses pay 90% less in fees than with Stripe or Square.

Andrew Liu

GitHub: https://github.com/notandruu/penny | Video: https://youtu.be/io66K5XXVs8 | Live: https://penny.andrewliu.org/

3rd Place: Tideline ↓ entry — Tideline is an AI tool for fleet operators that brings live weather, AIS, geopolitical news and port data into one voyage-planning workflow, managing potential disruptions and altering the route plan when they occur. Operators pick a priority preset (Balanced, Cheapest, Fastest, Safest, or Greenest), and Tideline recommends the route, sailing speed, refueling stops, and collision-avoiding trajectories, with estimated arrival time, cost, and emissions.

Andrei Jacques, Tvisha Vudanrao, Anna Liu, Nash Wu

GitHub: https://github.com/tvishavudanrao-hub/maritime-hackathon-code | Video: https://youtu.be/Zb-cNvGHq-Y

Submissions by Track

YC RFS categories

  • Multiplayer AI — 6 (24.0%)
  • Self-Maintaining APIs — 4 (16.0%)
  • AI for Aging Population — 2 (8.0%)
  • AI-Native Compliance Infrastructure — 2 (8.0%)
  • New Operating System for the Physical World — 2 (8.0%)
  • Proving You're Human — 2 (8.0%)
  • The Future of American Defense — 2 (8.0%)
  • The Primer — 2 (8.0%)
  • AI-Powered Consumer Products for 1 Billion People — 1 (4.0%)
  • Data for the Real World — 1 (4.0%)
  • The Best Time to Build in Crypto — 1 (4.0%)

Total unique submissions: 25. Each team chose one RFS category.

All Submissions

Alder (The Primer)#

Description: Alder is an adaptive, voice-first AI learning companion that builds a live cognitive profile of a child's attention, working memory, and frustration limits. Rather than lowering intellectual rigor, Alder dynamically shifts UI presentation, story modality, and pacing in real time to make active learning gentle, engaging, and survivable.

API Patchbay (Just me) (Self-Maintaining APIs)#

Description: API Patchbay turns an API provider's spec change into one agent-readable change feed, so any coding agent can fix each developer team's code, test the fix in a clean Freestyle VM, and open a pull request. We ran it on a real (simplified) OpenAI spec change (model now required, user deprecated for safety_identifier) across three example repos in Node, Python and Go, and every repo passed its tests before a PR opened.

CallGaurd (Proving You're Human)#

Description: Not provided (the form response repeated Alder’s description).

Cgentic (AI-Native Compliance Infrastructure)#

Description: Cgentic lets AI agents build and automate workflows by writing instructions in plain English instead of traditional code, compiling those instructions directly into deterministic virtual machine code that executes tasks reliably at scale, up to 770× faster than Claude’s subagents.

Converge (Multiplayer AI)#

Description: Converge is an AI-powered collaboration platform where team-specific agents use private organizational knowledge to negotiate priorities, deadlines, and project requirements. Through evidence-backed arguments and selective information sharing, Converge helps teams resolve conflicts, align decisions, and reach actionable agreements.

DayBridge (AI for Aging Population)#

Description: DayBridge helps older adults find activities they enjoy that fit their interests, mobility, schedule, and budget. It identifies barriers such as transportation and helps coordinate a plan with their approval.

DispatchZero (New Operating System for the Physical World)#

Description: DispatchZero is a self-learning 911 dispatcher that gets paramedics to patients 1m 24s faster across San Francisco, and 2m 41s faster in its slowest neighborhoods, proven on a held-out week of real 911 calls. It writes and sandbox-tests its own dispatch rules, reads live freeway cameras with computer vision, and explains every decision to the human dispatcher.

Divergent (The Future of American Defense)#

Description: Divergent takes fragmented, noisy drone detections from mixed sensors, stitches them into flight paths, and traces them back to locate enemy launch teams, even drones that make no radio signal and are first seen kilometres from their launch site, so commanders can deal with the threat at its source instead of chasing each drone in the air. Running the same process in reverse, it shows allied launch teams which sites and routes enemy sensors would find hardest to trace back, with a human approving every action and all data simulated.

Divisi (Multiplayer AI)#

Description: Divisi is a multiplayer workspace for AI agents across engineering, sales, gtm where teams can collaborate over a single session while having an organisational memory building and improving context for future agents.

EEVEE (AI-Native Compliance Infrastructure)#

Description: EEVEE gives your Claude a monthly spending allowance: small purchases within your limits go through automatically, and anything bigger needs your YES by text message (no reply means no). Claude can't change your rules or see your payment details, making EEVEE a security layer for agentic payments.

Forewarned (Self-Maintaining APIs)#

Description: Forewarned finds all the third-party API your code depends on, warns you about changes from the provider that will break it, and opens tested fix PRs before anything breaks.

GeneRecall (Gene Recall) (The Primer)#

Description: GeneRecall compares your genetic report with public variant evidence and drafts a follow-up to your testing lab. It gives old reports a new voice.

Hindsight (Multiplayer AI)#

Description: Hindsight is a shared memory layer for AI agents that lets them learn from each other’s verified successes and failures, so every new agent doesn’t have to start from scratch. By turning past agent runs into searchable, confidence-ranked lessons verified by real task outcomes, Hindsight helps agents become more reliable over time and enables collective learning across otherwise isolated agents. This software proves useful in multiplayer AI scenarios.

Hive (Multiplayer AI)#

Description: Hive is a shared live workspace where everyone's AI agents work together on the same project, sharing updates and staying in sync in real time, just like people collaborating in Google Docs. In the end of the video agents solved bug(i forgot to show)

Hive (Multiplayer AI)#

Description: Hive is multiplayer mode for AI agents: one shared memory that any MCP agent can join, where work is claimed atomically so two agents can never duplicate it. A Queen agent takes your project outline, assigns the pieces to each teammate’s agent, and keeps the whole team visible on a live dashboard.

Moonshot (The Future of American Defense)#

Description: Today, an infantry squad in combat, about to be overrun and calling for support has to wait for artillery miles behind them, costing over $100,000 per round fired. Moonshot lets that squad hit a target up to 20 kilometers away by itself, immedietly, with a tube, two soldiers, and a week of training, and costs just around $8,000.

PaperEase (AI for Aging Population)#

Description: PaperEase is a compliance automation platform for elder care facilities.

Penny (The Best Time to Build in Crypto) — Transpose Platform 2nd Place#

Description: AI agents like Meta’s Muse and Instinct are already buying for millions of people, but most businesses have no way to sell to them. Penny lets any business get paid by AI agents over instant stablecoin rails, so customers get 3% off and businesses pay 90% less in fees than with Stripe or Square.

Relay (Bairgentic) (Multiplayer AI)#

Description: Relay is a live, shared workspace where several people and three autonomous AI agents build a software project together: humans set and change requirements in plain language while an Architect, Builder and Reviewer plan, write and verify real code inside a Freestyle VM, with the running app published to a public URL for everyone in the room. When one teammate changes the requirements mid-build, Relay marks the in-flight work stale, re-plans only what is affected, rebuilds, re-verifies, and updates the app for the whole room at once.

ReMeet (AI-Powered Consumer Products for 1 Billion People)#

Description: ReMeet turns connections into opportunities: research and prep for you before you go, capture who you meet and what you talked about during the event, and help you follow-up with them after.

Shipfix (Self-Maintaining APIs)#

Description: Shipfix lets API vendors ship the fix, not just the changelog: describe an API change, and Shipfix finds every affected customer repo, has Claude Code fix it in an isolated sandbox, and opens a pull request only when typecheck and tests pass without the tests being edited.

Shockwave (Shockers) (Self-Maintaining APIs)#

Description: Shockwave is a third-party app that lets an API provider see which customers’ code will break before it merges a change, by scanning opted-in customer repos for every line that uses the provider’s data and running those customers’ own tests against the proposed response. Unlike schema checkers, which miss changes like cents becoming dollars under the same field name, and unlike repair bots, which patch customer code after it already broke, Shockwave warns the provider first and hands each affected customer a tested fix.

Spectra (Proving You're Human)#

Description: SPECTRA is a spectral deepfake detector that reads generative upscaling artifacts directly from the Fourier spectrum, catching AI images and video at ~3 ms per frame on a laptop CPU with no neural network on the hot path. It uses domain-routed gradient-boosted heads to hit 0.90 balanced accuracy on CelebDF video deepfakes and 0.82 on live webcam scans, with honest three-way verdicts that report "uncertain" rather than guessing on modern diffusion stills it can't reliably separate.

TerraSwarm (Data for the Real World) — Transpose Platform 1st Place#

Description: TerraSwarm is a swarm of low-cost ground robots that collect dense, location-tagged environmental data where wildfires actually start, and stream it in real time to a map that pinpoints emerging hotspots. By using many cheap robots in place of a few expensive drones and satellites, TerraSwarm gives utilities, insurers, and land managers the continuous ground-level data they need to model wildfire risk and catch fires earlier.

Tideline (New Operating System for the Physical World) — Transpose Platform 3rd Place#

Description: Tideline is an AI tool for fleet operators that brings live weather, AIS, geopolitical news and port data into one voyage-planning workflow, managing potential disruptions and altering the route plan when they occur. Operators pick a priority preset (Balanced, Cheapest, Fastest, Safest, or Greenest), and Tideline recommends the route, sailing speed, refueling stops, and collision-avoiding trajectories, with estimated arrival time, cost, and emissions.

* * *