c0mpiled_
← back to articles
hack·october 16, 2026·36 min read

YC: Own Your Intelligence Hackathon - September 2026

by neel majmudar273 reads

Overview

  • 200+ participants, 91 submissions (91 unique projects), 16 prizes awarded across 14 projects.
  • Hosted by River AI, GBrain, Memorable, QM, Superset, and UFO, with YC.
  • Theme: Own your intelligence: : memory, models, and harnesses that you run and keep, instead of renting them.

Reflections

Good

  • Strong sponsor-stack usage: 90 of 91 submissions tagged at least one side quest, and GBrain alone appeared in 65 (71.4%).
  • Many teams went past prompting and actually trained models they own, River AI fine-tunes showed up in 42 submissions (46.2%), several with held-out evals.
  • Quality of the projects were remarkable, majority of the projects displayed strong technical expertise.

Bad

  • Judging took a lot longer than what was scheduled.

Sponsors & Judges

91 submissions, 16 prizes awarded! Hosted by River AI, GBrain, Memorable, QM, Superset, and UFO.

Judges

Track assignments as provided; unallocated judges have not yet been assigned to a track.

River AI

UFO

GBrain

Superset

Memorable

YC

Others

Prizes & Recognition

Grand Prize

1st Place: QM Motion — QM Motion gives an AI coding agent a sense of time, so it can fix UI bugs that only show up in motion — janky animations, layout shifts, and elements that jump or snap back. In a sandboxed headless Chromium, its motion tool reproduces the interaction and records a short timestamped capture, and the agent inspects key frames itself to judge what went wrong. After editing, it records a fresh take, aligns it with the original at the moment of the click, and uses pixel diffs to show exactly what changed; each case is saved to GBrain so the agent recalls similar bugs next time.

2nd Place: 49th Labs (The 49th Engineer) — Your team's 49th engineer, trained on your other 48. The 49th Engineer is an AI engineer for hardware teams whose weights change directly from senior engineers' reviewed work: it watches CAD changes and Slack messages, stores every verified engineering change in GBrain with provenance, and fine-tunes a personal LoRA on River. Lessons that pass review, scope and regression tests are promoted into the shared company model through "weight pull requests" — a junior's conflicting rule was blocked in 1.8 s, while a senior's USB-A rule took the model from 0/12 to 12/12. On 30 held-out engineering changes, Qwen3.5-9B went from 0/30 to 29/30 after the River pipeline, and the whole thing runs as a QM agent in Slack.

3rd Place: Support AGI — A customer-support agent that gets cheaper with every ticket. It remembers how it solved each kind of request, replays that path next time, and compiles repeated ones into plans that run with zero model calls. Across 400 real support tickets, cost per ticket fell 66%.

Honorable Mention: GM — GM compiles Garry Tan's GBrain into an open-weight model you own. Today an agent re-reads about 5,000 tokens of GBrain rules on every call; GM learns all 75 skills into its weights, so it routes, files and cites with a 45-token prompt. Every night, the GM Nightly Loop trains it on your team's approved notes and corrections, and the new model goes live only if it passes a held-out gate.

Honorable Mention: goodnight (Countinghouse) — An open-source AI CFO for startups that are shipping and don't have time to tally financials. Countinghouse pulls Mercury, Brex, Ramp, Stripe, Gusto, AWS and Carta into one ledger and shows it as an Accounting tab inside QM: cash, burn, runway, a tax calendar, and a CFO brief of what changed and what to decide. An advisor agent suggests what to cut, renegotiate or build yourself, the QM agent answers finance questions and runs what-ifs, and GBrain keeps everything it learns as plain markdown.

Side Quests

Superset** — 1st:** Swarm Solver — Agent swarms with SMT solvers and a fast message server. Planned and designed with GStack.

Memorable** — 1st:** Fleetbrain — Fleet memory and repair intelligence for the upkeep of robots.

Memorable** — 2nd:** Waymark — Record once, guide everyone. Waymark is a Chrome extension that records how someone does a task on any website; Claude removes the wrong turns, and the workflow is saved to GBrain and Memorable and shared in QM. Teammates ask QM "how do I…?" and get a glowing, step-by-step guide in their own browser — it never clicks for you and never records what you type.

Memorable** — 3rd:** RoboDojo — RoboDojo is a sparring ground for agents to learn physical tasks — and every robot that trains there makes the next one smarter. Built on GBrain, Memorable, and QM.

UFO** — 1st (tie):** OYI: Own Your Inventory — An SMS/WhatsApp extension for UFO so non-technical people can use it from their phones, plus an extension that makes UFO useful for restaurant owners — inventory management and similar day-to-day tasks.

UFO** — 1st (tie):** Ufo + Memorable — A Memorable extension for UFO.

QM — 1st: Air_RSI — Continual learning in QM: a plugin that wires in models that continuously fine-tune themselves from QM conversations. It wraps a student/teacher model behind an inference API with A/B sampling, and in the demo trains a model to be proactive in workspaces from teammates' real reactions to its proactive messages — all inference and training on River.

GBrain** — 1st:** goodnight (Countinghouse) — also an honorable mention for the grand prize (see above).

River AI** — 1st:** voice — Train a small model to actually write like you — fine-tuned quickly and easily on River.

River AI** — 2nd:** GM — also an honorable mention for the grand prize (see above).

River AI** — 3rd:** AskJeeves — Digital twins on your team.

Submissions by Side Quest

Total unique submissions: 91. Many teams tagged multiple side quests, so percentages don't sum to 100%; one submission tagged none.

All Submissions

49th Labs (The 49th Engineer) (River AI · GBrain · QM · Superset) — 2nd Place#

  • Description: Your team's 49th engineer, trained on your other 48. The 49th Engineer is an AI engineer for hardware teams whose weights change directly from senior engineers' reviewed work: it watches CAD changes and Slack messages, stores every verified engineering change in GBrain with provenance, and fine-tunes a personal LoRA on River. Lessons that pass review, scope and regression tests are promoted into the shared company model through "weight pull requests" — a junior's conflicting rule was blocked in 1.8 s, while a senior's USB-A rule took the model from 0/12 to 12/12. On 30 held-out engineering changes, Qwen3.5-9B went from 0/30 to 29/30 after the River pipeline, and the whole thing runs as a QM agent in Slack.
  • Team: Omar Khalil
  • GitHub: https://github.com/OKKHALIL3/49th-labs
  • Video: https://youtu.be/KKwcHy36oT4

AgentPalace (River AI · GBrain · QM · Superset)#

AI Deal Maker (River AI · Memorable)#

Air_RSI (River AI · QM) — QM 1st#

Appstract (GBrain · QM)#

AskJeeves (River AI · GBrain · UFO) — River AI 3rd#

Barrier (River AI · GBrain · Memorable · QM · Superset · UFO)#

BiblioTech Health (GBrain · QM)#

Capture the Brain (GBrain · Memorable · QM · UFO)#

Certified Brain (River AI · GBrain · QM)#

  • Description: Long-term AI memory depends on cached judgments, and every prompt or model update forces a choice between rebuilding the cache or serving stale context. Certified Brain re-checks a small sample of each group of cached answers and re-evaluates only the groups that fail. On GBrain's 4,863-verdict contradiction cache, it used 1,267 calls with 48 wrong verdicts (versus 576 for blindly keeping the cache), staying within a 10% error budget in all 1,000 reruns.
  • Team: Arjun Lohan
  • GitHub: https://github.com/arjunlohan/certified-brain
  • Video: https://www.loom.com/share/1b05d52b7d7f42bdb7107a2715faca00

Chief of Staff (GBrain · QM)#

Clone-in-the-Loop (GBrain · QM)#

  • Description: Clone-in-the-Loop replaces human-in-the-loop with your clone. Built as a QM fork, it predicts what you'd ask your agent next, using GBrain to retrieve instructions and feedback from your imported Codex and Claude conversations. Tab once to accept a suggested prompt, twice to enable Clone mode — QM executes while your clone reviews results and requests corrections — and team workspaces let you guide work from a teammate's clone.
  • Team: Minchan Kim
  • GitHub: https://github.com/cloneismin/clone-in-the-loop
  • Video: https://www.loom.com/share/f989c778d4f14d05acfe0a20d9bdfbc3

Cole (Memorable)#

Command & Control (River AI · GBrain · QM · Superset)#

CompanyOS Cortex (River AI · GBrain · QM)#

cull (GBrain)#

  • Description: cull reads the morning firehose — 100+ new items from Hacker News and arXiv — and decides which 5 are worth your time. Every judgment lives in plain markdown you can open and edit, and thumbs up or down rewrites the taste memory so the next run is visibly sharper. One Python file, zero dependencies, answers in under a second.
  • Team: Ethan Lam
  • GitHub: https://github.com/ethanmlam/cull
  • Video: https://youtu.be/yqZ3C_zQoUw?is=Dk-wbIoLWHZwgJuU

Custom Security Model (River AI · Superset)#

Darwin Arena (River AI · GBrain · QM)#

Deltable (Memorable · Superset)#

Editude (QM)#

Faceplate (River AI · GBrain · QM)#

  • Description: An AI copilot for industrial control screens that learns your team's standards. Engineers click any part of an Ignition plant screen and say what to change in plain English; Faceplate pulls rules, equipment history and past decisions from GBrain, and a Qwen model trained on River makes the change in the team's ISA-101 house style — passing 14 of 20 unseen test edits where plain Qwen passed 0. Every accepted change is logged back to GBrain as training data, and 15 Ignition skills let the team's QM agents operate the plant the same way.
  • Team: Muhammad Zubair Zafar, Ali Amjad, Ayush Ojhah
  • GitHub: https://github.com/zubair480/bring-your-own-intelligence-yc-26
  • Video: https://faceplate-demo.pages.dev

FIA — Fraud Intelligence Agency (River AI · GBrain · Memorable · QM · Superset · UFO)#

FlashCode (GBrain · QM · Superset)#

Fleetbrain (GBrain · Memorable · QM · Superset) — Memorable 1st#

Floodgate (River AI · GBrain)#

  • Description: A Chrome extension that blocks pages that distract from the task you're actually on — the same video passes or fails depending on your work. The team rebuilt Open-Jev on River and trained a calibrated decision model (held-out accuracy 69.7% → 86.2%), plus a per-user model that reached 80.0% on one teammate's own browsing. GBrain remembers your corrections and Memorable recalls the procedure you're mid-way through.
  • Team: Phu Quach, Chinat Yu
  • GitHub: https://github.com/cyu60/floodgate
  • Video: https://youtu.be/HAqmeRNOy6g

four1 (GBrain)#

G-UI (GBrain · QM · Superset)#

GM (River AI · GBrain · QM · Superset) — Honorable Mention · River AI 2nd#

goodnight (Countinghouse) (GBrain · QM) — Honorable Mention · GBrain 1st#

  • Description: An open-source AI CFO for startups that are shipping and don't have time to tally financials. Countinghouse pulls Mercury, Brex, Ramp, Stripe, Gusto, AWS and Carta into one ledger and shows it as an Accounting tab inside QM: cash, burn, runway, a tax calendar, and a CFO brief of what changed and what to decide. An advisor agent suggests what to cut, renegotiate or build yourself, the QM agent answers finance questions and runs what-ifs, and GBrain keeps everything it learns as plain markdown.
  • Team: Charles Zheng
  • GitHub: https://github.com/goodnight000/countinghouse
  • Video: https://youtu.be/lDp7q1gS5Qw

Harmony (GBrain · Memorable · Superset)#

Host AI Engineer (River AI · GBrain · Memorable)#

Hot Mic (GBrain)#

Hydroxyl BaC (Superset)#

IMMUNE (River AI · GBrain · Memorable)#

Impactor (River AI · GBrain · Memorable)#

  • Description: An interactive 3D simulation of agents deciding which past results still apply when their world changes. Delivery robots face rain, heavy cargo, low battery and deadlines; experiments in GBrain carry conditions and outcomes, execution branches around Memorable procedures, Jev picks the strategy, and a River-tuned adapter chooses corrections from simulated preflight checks.
  • Team: Cosmin Bararu
  • GitHub: https://github.com/Cosmin-B/impactor
  • Video: https://goofy-chive-qlt6e.shipped.run/

Jelly (River AI · GBrain · Memorable · Superset)#

  • Description: Jelly sits between your coding agent and its model and turns the work your agents repeat into a model you own. It records every session and runs the task's own tests; Memorable sorts each run into a task type, and after 5 verified passing runs Jelly trains a small model (Qwen3.5-9B on River) and routes that type to it, falling back to the big model on failure. In the live run, River trained the model in 96 seconds, and it then fixed unseen tasks for 24–39× less per task.
  • Team: Ayaan Gazali, Jerry Xiao
  • GitHub: https://github.com/ayaangazali/jelly
  • Video: https://vimeo.com/1230759965

Knack Pack (River AI · GBrain · Memorable · QM)#

  • Description: Knack helps robots learn from their mistakes — and keep the lesson. The prototype focuses on a simulated robot arm that misses a grasp, using River AI fine-tunes for failure diagnosis, QM for recovery orchestration, Memorable for edge cases and procedural recall, and GBrain for knowledge and evidence. The goal is to recognize the same failure in a fresh run and recover without being taught again.
  • Team: Gursharan Singh
  • GitHub: https://github.com/guru3s/Knack
  • Video: https://youtu.be/Af5PbITse-s

kompX (River AI · GBrain · Memorable · QM · Superset)#

Lantern (River AI · GBrain · Memorable · QM · UFO)#

  • Description: A website help assistant that shows you the real button on your screen instead of giving you a paragraph. It answers from the company's brain in GBrain, says honestly when a feature doesn't exist and drafts a PR to build it for discussion in a QM room, and Memorable remembers routes customers finish. A model trained on River picks the right next button 77% of the time on new tasks, versus 40% for the base OpenAI model.
  • Team: Aadi Dahake
  • GitHub: https://github.com/AadiDahake/lantern-own-your-intelligence
  • Video: https://cap.so/s/vzdfd3f0ysp87wx

loopfour (River AI · GBrain · QM)#

LudoMachine (GBrain · QM)#

Manta Labs (River AI)#

Media Guard (River AI)#

Membrane (GBrain)#

Memento (River AI · GBrain)#

Mirror (River AI · GBrain · QM · Superset · UFO)#

Nailed It (River AI · Memorable)#

  • Description: A live multiplayer party game. A model reads each player from data on their own machine and makes assumptions about them; in a shared room, one player sits in the hot seat, a read appears, everyone guesses whether the model was right, and the hot seat reveals the truth. Every reveal becomes a training label for the reader.
  • Team: Iana Lin
  • GitHub: https://github.com/ianalin123/nailed-it
  • Video: https://ianalin123.github.io/nailed-it/

Nelly (River AI · Memorable)#

Ocean: Personal Sous Chef (River AI · GBrain · UFO)#

  • Description: A robot personal chef that learns how your house likes it done. A Franka Panda arm in a MuJoCo kitchen cooks a full breakfast in 15 steps, triggered over MCP from UFO, with every run and lesson stored as linked pages in a local GBrain. It recalls house habits before each step (the recipe says oil, this family uses butter), adapts mid-cook when you say "out of salt," and turns failed steps into lessons so it gets them right the next morning.
  • Team: Ocean
  • GitHub: https://github.com/r-agni/ocean
  • Video: https://drive.google.com/drive/folders/1Ht_4JJdHokp3rE5JVrYlq4SY65Pc3ZxC?usp=sharing

OpenSymbiote (River AI · GBrain · Memorable · QM · Superset)#

Over the Shoulder (GBrain · Memorable · QM)#

OYI: Own Your Inventory (UFO) — UFO 1st (tie)#

PacBrain (River AI · GBrain · Memorable · QM · Superset)#

PaymentOpsAgents (QM)#

Permissions for GBrain (GBrain · QM)#

present. (Memorable · QM)#

  • Description: A reality layer for personal AGI: your AI shouldn't stop knowing you when you close your laptop. On a Meta Quest 3S, the wearer just talks to people; OpenCV recognizes only opted-in people locally, Whisper transcribes on device, and Claude turns conversations into commitments, decisions and feature requests streamed into GBrain as person, relationship and event pages. A HUD anchors a person card on each face, and a QM fork makes reality a first-class input.
  • Team: Matthew Kim, Stephen Hung
  • GitHub: https://github.com/MatthewKim323/present
  • Video: https://youtu.be/7yvdC6-BHUo

Proceedingly (River AI · GBrain · Memorable · Superset)#

QM Edge (QM)#

QM Missions (QM)#

![][image6]

QM Motion (QM) — 1st Place#

  • Description: QM Motion gives an AI coding agent a sense of time, so it can fix UI bugs that only show up in motion — janky animations, layout shifts, and elements that jump or snap back. In a sandboxed headless Chromium, its motion tool reproduces the interaction and records a short timestamped capture, and the agent inspects key frames itself to judge what went wrong. After editing, it records a fresh take, aligns it with the original at the moment of the click, and uses pixel diffs to show exactly what changed; each case is saved to GBrain so the agent recalls similar bugs next time.
  • Team: Charles Conner
  • GitHub: https://github.com/charliec2004/qm-motion
  • Video: https://www.youtube.com/watch?v=LHMX7NuY_ao

QM Raid (River AI · GBrain · QM)#

  • Description: QM Raid turns a crew of AI coding agents into a real-time strategy army. Open issues are bug nests on an isometric map and every unit is a real QM agent; right-click a camp and the unit marches over, recalls what the team knows from GBrain (the Library), does the work in QM, and writes what it learned back. Units form teams with planner, implementer and reviewer loops, and Autopilot proposes orders with a 15-second veto.
  • Team: Emre Barut
  • GitHub: https://github.com/aemrebarut/qm-raid
  • Video: https://youtu.be/byesZLRnYho

QM TPM (QM)#

Receipts (GBrain · Superset)#

  • Description: Receipts holds public figures to what they said. Follow anyone and it finds their recent podcasts and interviews, keeps only claims it can match word for word to the transcript, and turns predictions into dated bets in your own GBrain. When a deadline passes, two AI judges grade it against web evidence, and you can ask questions like "Who has been most wrong about robotaxis?" and get answers grounded in dated receipts.
  • Team: Dhyey Mavani
  • GitHub: https://github.com/DhyeyMavani2003/receipts
  • Video: https://youtu.be/PMrdT9YDtGc

Red Team Range (QM)#

Reflex (River AI · UFO)#

Rema (GBrain · Memorable · Superset)#

Remnant (River AI · GBrain)#

  • Description: Remnant builds a reflection of someone you lost from their messages and a voice sample. Memories live in GBrain on your machine, the model is a River LoRA whose weights are yours, and the voice is cloned behind a consent gate. It never claims to be alive, never initiates, and steps aside entirely on crisis language. Built solo.
  • Team: Kartik Kapoor
  • GitHub: https://github.com/Kartikkapoor8/remnant
  • Video: https://youtu.be/H3T4W1cf5f4

RoboDojo (GBrain · Memorable · QM) — Memorable 3rd#

Rote (GBrain · Memorable · QM · Superset)#

  • Description: Rote turns a website into an API by watching an agent use it once. As a Claude Code agent drives Chrome, Rote records every request, templates the one that carried the answer, and verifies it with a probe word the agent never typed — then rote call answers new searches with one HTTP request (Hacker News: 35.0 s agent vs 128 ms call). An LLM judge checks the agent's claims against the recording, and GBrain stores each learned API and its proof.
  • Team: Patrick Liu
  • GitHub: https://github.com/machmoon/pilot-yc
  • Video: https://youtu.be/h92qHm0e8PM

Scout Autopilot (GBrain · Memorable)#

Scrubline (River AI · GBrain)#

  • Description: A launch weather officer: pick a rocket and a launch time and get a NO-GO to GO call with evidence and better windows. Weather caused 91 of the 154 Cape scrubs since 2020 with a known cause; Scrubline estimates all ten lightning launch rules, scores risk with a LightGBM model (ROC-AUC 0.735 on unseen 2025–26 launches), and gets a second opinion from Qwen3.5-9B fine-tuned on River, searching GBrain memory of 389 past missions for precedent. Moving just the riskiest 10% of launch hours would have avoided about 47% of weather scrubs.
  • Team: Saahith Veeramaneni, Jerry Chang
  • GitHub: https://github.com/SaahithV6/YCOYI9_27
  • Video: https://youtu.be/qbZB2rk05XE

Sombrero (Memorable · QM)#

Sophron (River AI · GBrain · Memorable · Superset)#

Spyglass (QM)#

Support AGI (River AI · GBrain · Memorable · QM) — 3rd Place#

Swarm Solver (Superset) — Superset 1st#

Swarm View (GBrain · QM)#

Swarmem (Memorable · QM · Superset)#

Talos (GBrain · Memorable · Superset)#

The Wall (River AI · GBrain · Memorable · QM · Superset)#

Throne (GBrain · QM)#

Together.AI (GBrain · Memorable)#

trycopilot.ai (GBrain · QM)#

Twin Gate (GBrain · QM)#

  • Description: Twin Gate puts a gate in front of agent writes to GBrain. An agent asked to change the brain works in a twin — a branch of the brain plus a forked, internet-isolated CreateOS sandbox — where five checks run (front matter, links resolve, copied figures match sources, no new outbound URLs, no key material) and every action lands on a hash-chained log. The owner sees the diff, checks and log on one page, and approve is the merge.
  • Team: Naman
  • GitHub: https://github.com/naman485/twin-gate
  • Video: https://drive.google.com/file/d/1kJGFzJJZvaNkTy_lnLXVMe7psQiBdYRb/view?usp=sharing

Ufo + Memorable (GBrain · UFO) — UFO 1st (tie)#

Undo for AI (GBrain)#

Viewfy Swarm (River AI · GBrain)#

voice (River AI) — River AI 1st#

Waymark (GBrain · Memorable · QM) — Memorable 2nd#

Which Childhood Is Yours (Souling) (GBrain)#

WorkflowDNA#

* * *