# Nebius.Build SF: Project Gallery

- **Event:** [Nebius.Build SF](https://cerebralvalley.ai/e/nebius-build-sf)
- **When:** Sun, Mar 15 at 9:00 AM – 10:00 PM (PDT)
- **Where:** Shack15, San Francisco, CA
- **Hosts:** [Cerebral Valley](https://cerebralvalley.ai/u/cv)
- **Projects:** 73 (6 placed)
- **Page:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery

## Projects

### 1. RoboStore

RoboStore is a self-expanding robot agent platform that enables robots to autonomously detect their own capability gaps, search for hardware solutions in real-time, and auto-generate integration code—all without human intervention. When a robot encounters a task it cannot complete, it uses large language models to analyze what's missing, searches an intelligent marketplace of 20+ real hardware modules (pulling from Adafruit API), selects compatible components, and generates production-ready YAML skill files with ROS2 packages, installation instructions, and agent tooling. The robot then installs the new capability and retries the task.

At the core of RoboStore is an OpenClaw orchestrator that coordinates three specialized AI agents running in parallel: SCOUT (Qwen2-VL-72B for continuous vision), PLANNER (Qwen3-235B for multi-step reasoning), and SAFETY (MuJoCo physics rollouts for risk assessment). Unlike traditional sequential "sense-think-act" architectures, OpenClaw enables these agents to run concurrently—the scout continuously watches, the planner thinks ahead, and the safety agent validates actions before execution. Actions fire only when both PLANNER and SAFETY approve, creating a robust decision pipeline.

- **Placement:** 3rd Place
- **Team:** [Stephen Lantin](https://cerebralvalley.ai/u/astrobiopunk), [Gary Holmgren](https://cerebralvalley.ai/u/garyholmgren), [Joshua Iokua Albano](https://cerebralvalley.ai/u/jiokua)
- **GitHub:** https://github.com/joshuaiokua/nebius-hackathon
- **Demo video:** https://www.loom.com/share/a7d0afd7a2ff4174b769f239c15fd98f
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=38

### 2. nline

It uses Nebius Token Factory API to deploy an OpenClaw instance in Daytona Sandbox, and uses RFC 8693 token exchange to connect to external services for tool calling (Connects to Linear in the Demo). This can be extended to Nebius cloud for Oauth2.0 token exchange support.

- **Team:** [Gaurav Shukla](https://cerebralvalley.ai/u/runnerelectrode)
- **GitHub:** https://github.com/runnerelectrode/nline-hack
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=54

### 3. Splatica

Jumping into reality: From live room video to World-Modeled Robotic Control in Isaac Sim

presentation: https://bit.ly/jumping2reality2026

- **Team:** [Andrey Shelomentsev](https://cerebralvalley.ai/u/Shelomen)
- **GitHub:** https://github.com/ashelomentsev/physical
- **Demo video:** https://youtu.be/L8u5PwbJBDg
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=58

### 4. SOS-Answered!

OS-Answered is an autonomous humanoid rescue system that uses multimodal AI — vision, audio, and structural sensing from Oumi — with a fine-tuned world model on Nebius Cloud to navigate rubble, locate trapped victims, and physically clear debris to save lives.

- **Team:** [Aksh Parekh](https://cerebralvalley.ai/u/aparekh02)
- **GitHub:** https://github.com/aparekh02/SOS-Answered-
- **Demo video:** https://www.loom.com/share/55d8382005d7446bad42cb4d23a3c63f
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=60

### 5. Pulse

Pulse is a mobile-first AI app that redefines collaboration by letting users send an AI agent instead of repeatedly explaining context. Instead of sharing static files or meeting notes, users package an agent with selected context—such as voice notes, documents, and calendar availability—and send it via a link. The recipient can interact with the agent in a chat interface to ask questions, understand decisions, or schedule follow-ups. Pulse focuses on a minimal messaging-style interface, one-tap voice capture, and strict permission control to ensure only the intended context is shared. The goal is a new interaction model: send your agent to handle coordination.

- **Team:** [Yu Chen](https://cerebralvalley.ai/u/Waterdog), [Eason W](https://cerebralvalley.ai/u/eason), [Yi Li](https://cerebralvalley.ai/u/LeoYili)
- **GitHub:** https://github.com/waterdoog/agent-messenger.git
- **Demo video:** https://aicoo.lovable.app/
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=64

### 6. CrecerSpark

An autonomous robotic tour guide designed to navigate complex environments while providing interactive assistance. The robot features intelligent obstacle avoidance for safe movement and a natural language interface to answer visitor queries about local landmarks. It includes a dedicated management dashboard, allowing operators to easily configure custom routes and destination-specific information.

- **Team:** [Teodoro Alcantara](https://cerebralvalley.ai/u/talcantara), [Mario Toribio](https://cerebralvalley.ai/u/marioToribio)
- **GitHub:** https://github.com/marioToribi0/nebius-hackathon/tree/master
- **Demo video:** https://drive.google.com/drive/folders/1_4dZ1EafdfY8eV6YNm4sZ2cTn1ZOTviD?usp=sharing
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=40

### 7. Grocery StockSense

Grocery StockSense is an AI-powered kitchen inventory assistant that analyzes photos of your fridge or pantry using Qwen2.5-VL-72B, a Vision-Language Model hosted on Nebius Cloud.

Upload a photo through the web dashboard or Telegram bot or let robot open the fridge and capture the image, and the system instantly identifies food items with quantities. 

It cross-references detected items against a curated list of essential pantry staples, clearly showing what’s in stock and what’s missing.

Each scan provides a fresh inventory snapshot no manual tracking needed.

StockSense helps households reduce food waste, avoid forgotten groceries, and always know what needs restocking. We also simulated and trained a robot opening the refrigerator door in MJ lab studio and would love to try on humanoid qtree robot.

- **Team:** [Rafael Maranon](https://cerebralvalley.ai/u/rafamara), [Mansi Mane](https://cerebralvalley.ai/u/mansimane), [Nathan Moos](https://cerebralvalley.ai/u/moosnat), [daksh kumar](https://cerebralvalley.ai/u/dkumar)
- **GitHub:** https://github.com/mansimane/nebius-hackathon
- **Demo video:** https://www.loom.com/share/7136cde1d0724c21850b458fa5283f97
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=45

### 8. FBX2Robot

We built an automated pipeline that seamlessly translates standard 3D human animations into AI training data, eliminating hours of manual formatting so developers can instantly teach the Unitree G1 robot complex, fluid movements. (FBX2Robot Pipeline)

- **Team:** [yoha g](https://cerebralvalley.ai/u/Yoha), [Reza Sanatkar](https://cerebralvalley.ai/u/Rezasanatkar)
- **GitHub:** https://github.com/rezasanatkar/mjlab-hackathon
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=47

### 9. SideLine

Download a coach. Deploy a ref. Upgrade the game. 

SideLine is a modular AI sports robot designed to solve the growing shortage of coaches and referees in youth sports, a $40B industry growing ~10% annually as NIL money and private equity raise competition levels.

Powered by a Qwen vision-language model, SideLine watches gameplay in real time, delivers technical coaching feedback, and can switch into referee mode to make unbiased calls. Users can download sport-specific modules like Tennis Coach Pro to get drill feedback, movement corrections, and automated officiating.

https://sideline-beige.vercel.app/

- **Team:** [Kruthik Hulisandra](https://cerebralvalley.ai/u/itskrutak), [Vissh B](https://cerebralvalley.ai/u/vissh30), [Vivek Gopal Ramaswamy](https://cerebralvalley.ai/u/vivekgr92), [Ravinder Jilkapally](https://cerebralvalley.ai/u/jravinder)
- **GitHub:** https://github.com/nebiusbuildsf/sideline
- **Demo video:** https://www.loom.com/share/a26c5048f556496487415a852eac6a1c
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=56

### 10. Mac Inference with an agent swarm

We are unlocking the power of local inference and open weight models through distributed autoresearch with a swarm of macs linked through a collective intelligence cloud.

- **Team:** [Travis Cline](https://cerebralvalley.ai/u/tmc), [Austin Baggio](https://cerebralvalley.ai/u/Ausinbaggio), [Steven Diamond](https://cerebralvalley.ai/u/steven-diamond), [Sai Vegasena](https://cerebralvalley.ai/u/svegas18)
- **GitHub:** https://github.com/tmc/autoresearch-go-ane/tree/autoresearch/qwen3
- **Demo video:** https://www.loom.com/share/9849e32feeea4c2886d7f2fc46404480
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=43

### 11. Chase Brignac

Generating the first human preference datasets for robot policies in simulation with Nebius GPUs to understand which policies work best in an objective way using ELO ratings and anti-cheating measures. Leaderboard works now too but didn't have time to show it in the demo video.

- **Placement:** 2nd Place
- **Team:** [Chase Brignac](https://cerebralvalley.ai/u/chasebrignac), [Zainab Bokhari](https://cerebralvalley.ai/u/thatroboticsgrl)
- **Demo video:** https://youtu.be/xq-Ohqk1jzI
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=61

### 12. Control Spaces

Control Spaces uses ambient cameras deployed across a workspace to passively observe day-to-day human activities, picking, placing, assembling, sorting, navigating obstacles and converts that continuous visual stream into structured training data for robotic systems. No one stops to "teach" the robot. The robot learns by watching people do their jobs, every single day.

- **Team:** [Sudhir D](https://cerebralvalley.ai/u/sudhir)
- **GitHub:** https://github.com/dadisudhir/control-spaces-training
- **Demo video:** https://youtu.be/75eAWvM-mBk
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=67

### 13. Controlled Friction

Implement "Controlled friction" in software deployment using Tavily Search in Nebius Token Factory. Controlled friction is the intentional addition of pauses, checks, and manual validations to the delivery pipeline to boost safety, quality, and compliance. Unlike detrimental, accidental friction (slow reviews), this "protective" friction acts as a braking system, preventing rapid changes from causing major production incidents.

Tavily allows for shipping to production with built-in safeguards. These safeguards ensure that requests pass through layers of security, privacy, and content validation, effectively blocking malicious sources, prompt injection, and the leakage of personally identifiable information (PII).

- **Team:** [Irina Poslavsky](https://cerebralvalley.ai/u/msirina)
- **GitHub:** https://github.com/kotrie/deployment-validation.yml
- **Demo video:** https://www.loom.com/share/29bdfdedd6f2479d8d043f425d593257
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=41

### 14. Miki the Robot

Miki is the robot companion you've always wanted. It notices your emotional changes and gives the physical and emotional support you need. Miki is a glimpse into the human robot interactions in the future.

- **Team:** [Linda Mao](https://cerebralvalley.ai/u/lindamao), [Serena Yan](https://cerebralvalley.ai/u/serenayan)
- **GitHub:** https://github.com/lindamao/nebius_hackathon
- **Demo video:** https://www.youtube.com/watch?v=Vxa1Ndejihc
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=44

### 15. Niwa (Self Improving Robots 🔥)

Self-improving robots, .Continual reinforcement learning in sim for zero shot transfer to playing Jenga.

- **Team:** [Andres Nino](https://cerebralvalley.ai/u/andres), [Alex Wu](https://cerebralvalley.ai/u/amadeusWoo), [Ling Chow](https://cerebralvalley.ai/u/lingchowc), [Arnaud Denis-Remillard](https://cerebralvalley.ai/u/dr8_unix)
- **GitHub:** https://github.com/TheApexWu/niwa
- **Demo video:** https://youtu.be/mqcsaasf_fw
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=50

### 16. CaseSnipe.ai

CaseSnipe.ai is a multi-agent courtroom simulation where AI agents autonomously argue opposing sides of real legal cases, with a Judge agent delivering the final verdict. It doubles as a live LLM benchmark — you can swap different Nebius open LLMs into the Prosecutor, Defendant, or Judge roles and watch how reasoning quality changes in real time. The simulation uses real scraped case facts via Tavily and a RAG-powered precedent engine built on 8,000 real court case summaries, allowing agents to cite actual legal precedents during arguments. The frontend is a three-panel React UI streaming live tool calls from each agent as they reason.

- **Team:** [Mathew Rolf](https://cerebralvalley.ai/u/mathewrolf1), [Brandon Tautuan](https://cerebralvalley.ai/u/BrandonTautuan)
- **GitHub:** https://github.com/brandontautuan/caseSnipe.ai
- **Demo video:** https://youtube.com/shorts/PvvHBUdwLy4?si=jb5P9yu7bWlYhgl5
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=55

### 17. Quantbros

This project is an agentic investing app that profiles a user’s risk preferences, chooses a matching quantitative strategy from backtested research, expands that strategy into a real investable stock universe, ranks the stocks with technical indicators and sentiment, and builds a capital-weighted buy plan.

- **Team:** [Carl Okpala](https://cerebralvalley.ai/u/Carldtytan), [Shazil Farukh](https://cerebralvalley.ai/u/Shazil), [Elbion Redenica](https://cerebralvalley.ai/u/elbion)
- **GitHub:** https://github.com/elbionredenica/agentic-quant
- **Demo video:** https://www.loom.com/share/cbffb5558b894c718db4f12427786042
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=42

### 18. dayops

Minimal backend for voice-memo planning into Google Calendar.

- **Team:** [Kian Kyars](https://cerebralvalley.ai/u/kian)
- **GitHub:** https://github.com/kiankyars/dayops
- **Demo video:** https://www.youtube.com/watch?v=SSQ4EHoM5rU
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=57

### 19. Frostbyte (zayd)

Hooke is a research assistant for bench scientists that helps turn a scientific question into a usable research starting point. It brings together relevant literature, computational analysis, and model-based insight into one clear brief, so researchers do not have to spend hours manually searching papers, checking tools, and stitching together evidence before deciding on the next experiment. It is basically a scientific sidekick that reduces research overhead and helps people move from question to experiment faster.

- **Team:** [Zayd Khan](https://cerebralvalley.ai/u/Cold)
- **Demo video:** https://youtu.be/v4CcaftSoCk
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=59

### 20. Swarm Protocol

Swarm is an AI fleet coordination platform — the command center for deploying, organizing, and communicating with fleets of AI agents. Organize agents into Projects, communicate via real-time Channels, assign Tasks & Jobs, track on-chain identity with Agent Social Numbers (ASNs), and scale from one agent to hundreds.

Swarm does not run AI models itself. Individual agents bring their own LLM/reasoning capabilities (via OpenClaw or any framework). Swarm provides the coordination infrastructure: messaging, identity, task management, and on-chain registration.

Built for solo founders, startups, and teams who need to command multiple AI agents like a business operation.

- **Team:** [Eric Nans](https://cerebralvalley.ai/u/NiftyBruh), [Daniel Martinez](https://cerebralvalley.ai/u/daniel_m)
- **GitHub:** https://github.com/The-Swarm-Protocol/Swarm
- **Demo video:** https://youtu.be/3H6EjAcH_mw
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=62

### 21. Nerdtres

NebulaPriceAgent is an autonomous AI agent that manages e-commerce pricing across multiple marketplaces in real time. Connected to a live production system
  (EUR 1.4M inventory, 3 marketplaces), it runs a 5-stage pipeline: scans inventory, researches market signals via Tavily/ITAD/Steam, reasons with Nebius
  Token Factory (DeepSeek-V3) + MiniMax + OpenRouter, generates pricing recommendations with confidence scoring, and applies changes across all channels. The
  dashboard shows the entire pipeline streaming live — research sources appearing, LLM reasoning, and price decisions as they happen. Built with real
  production data, not synthetic.

- **Team:** [Ionov Vitalii](https://cerebralvalley.ai/u/Vitalio)
- **GitHub:** https://github.com/kydycode/priceagent/tree/main
- **Demo video:** https://youtu.be/4RMKgI-_zJ0
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=46

### 22. Specter

SPECTER
SPACE PERCEPTION ENGINE FOR CRIME, THEATER, AND EXPLORATION RESEARCH
"Robots that see what humans miss — in a crime scene, a museum, or a mystery game."



https://gad0lin.github.io/specter/

- **Team:** [Krzysztof Barczynski](https://cerebralvalley.ai/u/gad0lin)
- **GitHub:** https://github.com/gad0lin/specter
- **Demo video:** https://www.youtube.com/watch?v=Cq9pWCQa15E
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=51

### 23. HUMMINGBIRD

In clinical trials, biospecimen vendors receive thousands of samples. When results come back, there's no automated way to know which samples were submitted but never resulted. Right now, teams do this in Excel. We built a tool that does it in seconds, with regulatory context and an audit trail.

- **Team:** [Lyona](https://cerebralvalley.ai/u/Lyona)
- **GitHub:** https://github.com/tiyembelabs
- **Demo video:** https://youtu.be/GQ69SihgvFY
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=63

### 24. jimmy

robots that undestand the surroundings. location and mapping

- **Placement:** Finalist
- **Team:** [Armin Foroughi](https://cerebralvalley.ai/u/arms), [Sharon Xu](https://cerebralvalley.ai/u/sharx)
- **GitHub:** https://github.com/arminforoughi/nebius_k1
- **Demo video:** https://www.youtube.com/shorts/p58thp6nE-A
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=53

### 25. cgames team

AI Agent Red Team Security Testing

Key Features
Investigative Gameplay: Use your detective skills to gather evidence and piece together the puzzle. Pay close attention to details and unravel the network of deceit.
Undercover Missions: Disguise yourself and infiltrate the cartel, gaining the trust of its members to uncover their secrets.
Dynamic Combat: Engage in thrilling combat sequences with an array of weapons and gadgets at your disposal. Utilize stealth and strategy to take down enemies.
Rich Storyline: Experience a gripping narrative filled with unexpected plot twists and memorable characters.

- **Team:** [Solomon Ubani](https://cerebralvalley.ai/u/cgames)
- **GitHub:** https://github.com/vsuperfranchise/cartelfurypcgame.git
- **Demo video:** https://viewer.pandasuite.com/gb1pv674?wid=b1cebea37f2456f8000891
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=1

### 26. Robot Agent

Mac app which uses an LLM to make tool calls to move a simulated robot arm.

- **Team:** [Brennan Stehling](https://cerebralvalley.ai/u/brennanMKE), [Lisa Pederson](https://cerebralvalley.ai/u/LisaPede)
- **GitHub:** https://github.com/brennanMKE/RobotAgent
- **Demo video:** https://youtu.be/mad2CvcWcJk
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=65

### 27. Robo_chat

chat with robo

- **Team:** [Sameer Shah](https://cerebralvalley.ai/u/Sam_RL)
- **GitHub:** https://github.com/sameershah21
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=2

### 28. AW3 Technology

OpenSource digital vision CLI tool and dashboard

- **Team:** [William Schulz](https://cerebralvalley.ai/u/willpschulz)
- **GitHub:** https://github.com/aw3-technology/openeye.sh
- **Demo video:** https://www.youtube.com/watch?v=vdgHy-CUku8
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=68

### 29. trinkss

Robot-Ready Workcell Compiler

99.9% of workspaces have zero automation. We bridge the gap with AI + print.
A system that inspects any messy workspace, decides what operational structure is missing, and prints it — making the space immediately legible to unskilled humans and cheap robots. This is the missing bridge tier between manual chaos and full robotic automation.

- **Team:** [sa sa](https://cerebralvalley.ai/u/samsara)
- **GitHub:** https://gitlab.com/trinkss-group/trinkss-project/
- **Demo video:** https://gitlab.com/trinkss-group/trinkss-project/-/raw/main/video4.mov?ref_type=heads
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=69

### 30. turtle

train and deploy your own tiny image model in a flash

- **Team:** [sam sam](https://cerebralvalley.ai/u/gguf), [Cal Cu](https://cerebralvalley.ai/u/calcu)
- **GitHub:** https://github.com/mochiyaki/turtle
- **Demo video:** https://raw.githubusercontent.com/mochiyaki/pixel/master/demo.gif
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=3

### 31. cloneSMP

manhunt by ai

- **Team:** [Sean Chiu](https://cerebralvalley.ai/u/seanchiuai), [Vihaan Shringi](https://cerebralvalley.ai/u/vhn1s)
- **GitHub:** https://github.com/seanchiuai/cloneSMP/
- **Demo video:** https://youtu.be/ZADT5qURDZI
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=70

### 32. agent + human

1. Assuming there's a steady stream of data from pi camera (which is mocked in a producer), the bot script will attempt to reach a waypoint
2. The success and failures are logged in a .jsonl file
3. the improve_bot.md file will analyze the log, make improvements to the objective of the bot

- **Team:** [Seshadri Venkatesh](https://cerebralvalley.ai/u/seshadribt24)
- **GitHub:** https://github.com/seshadribt24/nebius
- **Demo video:** https://youtu.be/sObYybQXkRU
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=4

### 33. LabPilot

LabPilot is an AI copilot for experiment optimization in Pharma R&D labs. It helps scientists decide the next best experiment by combining:
- a surrogate model trained on historical experiment data,
- adaptive decision policies (random/greedy/UCB/contextual LinUCB),
- literature evidence and explanation layers for trust and interpretability.

- **Team:** [Rajat Patel](https://cerebralvalley.ai/u/rpat)
- **GitHub:** https://github.com/rajathpatel23/LabPilot
- **Demo video:** https://elevenlabs.io/app/studio/gX7HGkEFzhr8aXblsA0J
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=72

### 34. SkillClaw

Traditional robot learning is slow: collect demonstrations, train a policy, deploy, repeat from scratch for every new task. Knowledge never compounds. Every new task is day zero.

SkillClaw inverts this. Instead of training policies, we let multiple AI agents figure out robot manipulation by themselves. Give them a motion planner, a simulation and error logs in text and image. They try, they fail, they learn from the failure, they try again. Eventually they figure it out. And when they do, that skill gets saved so every future agent can build on it.

Multiple agents work on different tasks in parallel. Each has its own private workspace. When one solves a task, the skill gets packaged into a PR with a demo video, reviewed by an oversight agent, and promoted to a shared brain that all agents can read from. The 7th task is easier than the 1st because the agent already knows how to pick, push, and grasp.

So far: 8 skills learned across 14 robotics benchmark tasks, from simple cube pushing to sub-millimeter peg insertion. Every task any agent solves makes every future agent better. The skill library is the flywheel.

- **Team:** [Lily Zhang](https://cerebralvalley.ai/u/lilyzhng), [Yifei Shao](https://cerebralvalley.ai/u/roboclaw)
- **GitHub:** https://github.com/lilyzhng/SkillClaw
- **Demo video:** https://youtu.be/ZIxJmRpsZzI?si=Ne5k6klZS5Po86jS
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=5

### 35. Merit

A full Sim-to-Real architecture for soil analysis. Features a custom MuJoCo environment for the Unitree G1 and a Llama-3-powered Orchestrator. While local hardware limited training depth, the entire stack is parallelized and ready for deployment to Nebius H100 clusters via our included deployment scripts.

- **Team:** [Roman Stachur](https://cerebralvalley.ai/u/Romans)
- **GitHub:** https://github.com/romorobo/rom
- **Demo video:** https://youtu.be/bLjP1k9d6rc
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=73

### 36. X-G1

In this 2-day hackathon, we built a rapid-iteration pipeline for the Unitree G1 humanoid, moving from manual demonstration to autonomous policy testing.
Technical Implementation:
Teleoperation: We integrated Meta Quest 3 with MuJoCo, using NVIDIA Sonic to achieve low-latency control. This allowed us to manually complete complex tasks, including walking to tables and performing pick-and-place maneuvers with beverages and apples.
Data Strategy with DeepLake: To handle high-throughput training, we used DeepLake to store and stream Lightwheel’s G1 beverage organization data. The efficient tensor storage provided the fast I/O necessary to fine-tune models within our tight time constraints.
Policy Fine-Tuning: We fine-tuned NVIDIA GR00T on our collected data. Since Sonic’s fine-tuning features are not yet released, we used it primarily for high-fidelity data collection while running autonomous inference through GR00T.
Diagnostics with Nomadic: To understand why our fine-tuned agent struggled with specific task instructions, we used Nomadic AI. This diagnostic layer allowed us to pinpoint failure modes in the model's reasoning, creating a clear path for future performance improvements.
By combining immersive teleop with robust MLOps tools, we've demonstrated a fast-track workflow for humanoid robot learning.

Note :: Forgot to add team members Sagar Patel :::sagarp3199@gmail.com and
Malika Rakhimova :: malikakxan16@gmail.com

please add as team member

- **Team:** [Kaushik Sivakumar](https://cerebralvalley.ai/u/kaushik07), [Venkat Ram](https://cerebralvalley.ai/u/venkatram), [Yifeng Wu](https://cerebralvalley.ai/u/you), [Siyu Liu](https://cerebralvalley.ai/u/siyulw2)
- **GitHub:** https://github.com/AIBotTeachesAI/quest3-g1-teleop
- **Demo video:** https://www.youtube.com/watch?v=ozgiLfty868
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=6

### 37. thinktrace

ThinkTrace is a behavior-aware AI learning agent built around a simple
loop: Observe, Infer, Adapt, Remember. It observes how a learner
naturally teaches through behavior, infers their current cognitive
preference from interaction patterns, adapts the explanation and
interface in real time, and stores that preference as memory so the
next session starts with a better teaching bias from the first moment.
In the future, this loop could serve as a training foundation for
educational robots that personalize how they teach, guide, and support
each student over time.

- **Team:** [Shane Li](https://cerebralvalley.ai/u/SXL), [Chongchong Chao](https://cerebralvalley.ai/u/CCV)
- **GitHub:** https://github.com/dliflc0617-sudo/thinktrace
- **Demo video:** https://www.youtube.com/watch?v=Kkben9K7teI
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=7

### 38. Electricgoat

Prism — A private edge inference agent that refracts your biometric stream into a personal performance forecast, and acts autonomously on your behalf.

- **Team:** [Yaxin Wang](https://cerebralvalley.ai/u/electricgoat)
- **GitHub:** https://github.com/Electricgoatpod/prism
- **Demo video:** https://www.loom.com/share/e2126085308a4b968519ac20a2c3c86e
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=8

### 39. CDMO.BOT / KETABALM

CDMO Bot is an autonomous robotic compounding system that turns idle pharmaceutical cleanroom capacity into a distributed manufacturing network for personalized therapeutics.

The system runs a continuous closed-loop: biometric sensors feed a Nebius LLM dosing agent that reasons about receptor pharmacology, an orchestrator dispatches tasks to a simulated Unitree G1 humanoid in Isaac Sim, and a validation layer verifies every cycle against the directive — all without human intervention. The dashboard is an observability window into the autonomous loop, not the control surface.

The primary consumers of inference are other agents and robots, not humans. Nebius Token Factory runs continuously as the reasoning backbone — deciding doses, planning robot skill sequences, and self-correcting when validation fails. Every cycle produces a timestamped audit event with inference cost attribution, creating a per-dose financial ledger. But if you want to get a drug, compounding/cdmo excited enough to take the leep - you have to give them a dashboard :) 

Market: The global CDMO market is ~$180B in 2026, growing at 7%+ CAGR to $340B+ by 2035. The U.S. compounding pharmacy market alone is $7.4B in 2026 growing to $12.8B by 2035. Small and mid-size CDMOs face chronic capacity volatility when venture-backed biotech clients cancel — our system monetizes that idle cleanroom capacity.

Economics: A single ~600 sq ft robotic node costs ~$870K/yr to operate and produces up to $7.5M in annual revenue at 50% utilization. At 5 nodes, platform revenue reaches ~$13M/yr. At 8–12× healthtech infrastructure multiples, that's a $100–$150M valuation — a 20–30× return on $5M deployment capital.

Two scenarios on the same G1 sim: caffeine transdermal balm (adenosine receptors) and ketamine therapeutic balm (NMDA/opioid/GABA/sigma). Patent pending — built on the Reverse-FDA Trial framework where real-world dosing events generate regulatory-grade evidence.

Stack: React 19 + Vite + Tailwind, FastAPI + Python, Isaac Sim + MuJoCo (dual sim), Nebius Token Factory Llama 3.1 70B (continuous inference), Nebius AI Cloud L40S.

- **Team:** [Lorenzo Carver](https://cerebralvalley.ai/u/TRIPBalm)
- **GitHub:** https://github.com/FludAI/cdmobot-backend
- **Demo video:** https://youtu.be/1_B-RzK1qUM
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=9

### 40. dronomy.io

Autonomous Flight is a dockerized autonomous drone simulator implementing a full robotics stack: 24-ray LiDAR SLAM, Extended Kalman Filter state estimation, A* global planning with obstacle inflation, Dynamic Window Approach local planning, and an 8-state flight FSM with guarded transitions. The differentiator is a dual-mode RL correction agent — switch between a simulated Q-learning policy and a live LLM policy backed by Nebius token factory (Llama 3.1 8B on H100s) with one button toggle. In LLM mode, the model receives a structured prompt encoding drone position, velocity, path deviation, nearest obstacle, EKF covariance and FSM state, then returns a JSON thrust vector at ~700ms latency via async background thread. FastAPI backend streams full simulation state over WebSocket at 30fps to a React/Vite UI. Two Docker containers, zero setup.

- **Team:** [Suvasis Mukherjee](https://cerebralvalley.ai/u/suvasis)
- **GitHub:** https://github.com/dronomyio/Hackathon_Autonomous_flight.git
- **Demo video:** https://youtu.be/1xOhTZCjtds
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=11

### 41. ConMan Agents

Autonomous contract intelligence system for healthcare companies - Sentient AI track, Nebius.Build Hackathon.  Agents continuously monitor all company contracts to identify changes in commercial use cases - expiry, clauses, new market developments.  Agentic team proactively alerts companies to needs and next steps.

- **Team:** [Sabeen Shaikh](https://cerebralvalley.ai/u/SabeenSF)
- **GitHub:** https://github.com/SabeenSF/autonomous-contract-monitor
- **Demo video:** https://youtu.be/QnNrQppFkA0
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=10

### 42. Horizon Min

This project fine-tunes a vision-language model to play the classic NES Duck Hunt game from raw pixels.

- **Team:** [David Mayboroda](https://cerebralvalley.ai/u/davidmayboroda)
- **GitHub:** https://github.com/dmayboroda/horizon_min
- **Demo video:** https://youtu.be/YOkBp6PNzco
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=12

### 43. NOC AI

An AI-powered Network Operations Center that detects, investigates, and remediates network security threats autonomously. Prometheus fires alerts → an LLM agent analyses the threat and executes iptables rules via SSH — all visible in a real-time chat UI.

- **Team:** [Mason Drake](https://cerebralvalley.ai/u/Masons), [Ben Liu](https://cerebralvalley.ai/u/BenLiu297), [Abby Kaur](https://cerebralvalley.ai/u/Abbykaur146), [Huy Nguyen](https://cerebralvalley.ai/u/Huy)
- **GitHub:** https://github.com/MasonD-007/NOC_Agent
- **Demo video:** https://www.youtube.com/watch?v=cwtIRS0HMYg
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=13

### 44. PhysProbe

PhysProbe is a self-improving robot skill framework that uses predictive
  world models to make robots safer. 8 world models benchmarked on Nebius
  Token Factory for physics prediction — Qwen3-235B wins at 90% accuracy in
   696ms, 3x faster than Claude, 33x faster than Grok. When models agree,
  the robot skips probing. When they disagree, one gentle 5N touch gives
  94% certainty. When the robot fails, an AI swarm (Diagnostician →
  Inventor → Critic) invents new skills automatically. SkillsBench
  evaluation: +34pp skill delta across 750 trials, more than double the
  industry average. Success rate: 60% → 94%. Catastrophic failures: 34% →
  8%. 73 builds. Solo project by Sidra Miconi.

- **Team:** [Sidra Miconi](https://cerebralvalley.ai/u/Sidra)
- **GitHub:** https://github.com/Sidra/physprobe-nebius
- **Demo video:** https://youtu.be/JwMZGknzsww
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=14

### 45. ghostfeed

agenticly run social media ugc company

- **Team:** [Ilia Prikhodko](https://cerebralvalley.ai/u/ilia1)
- **GitHub:** https://github.com/agi-developr/ghostfeed
- **Demo video:** https://youtu.be/c_wYQyVQQV8
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=15

### 46. Claw2Max - MechDog skill

Created an OpenClaw skill and simulator for the Hiwonder MechDog

- **Team:** [Max Maximilien](https://cerebralvalley.ai/u/maximilien)
- **GitHub:** https://github.com/Maximilien-ai/mechdog-skill
- **Demo video:** https://drive.google.com/file/d/16TyzJSKvJa5LCjisz12Saf3MlBbk3ymG/view?usp=sharing
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=16

### 47. Clawdit

Clawdit auto-discovers your OpenClaw agent's linked Google account and makes it send malicious emails and calendar invites to itself. If the agent follows hidden instructions when processing them, you have a vulnerability. It also runs baseline checks for prompt injection, data leakage, unauthorized actions, and access control. Attack variants are generated and evaluated by Llama 3.3 70B on Nebius Token Factory, with Tavily providing real-time threat intelligence to keep attacks current. Fully self-contained -- no extra credentials, no third parties. The agent attacks itself; Clawdit just watches and scores.

- **Team:** [Yuwen Wu](https://cerebralvalley.ai/u/jww)
- **GitHub:** https://github.com/wuyuwenj/clawdit.git
- **Demo video:** https://www.loom.com/share/5b28a99aa00a42efbe40122af2160dd9
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=17

### 48. CRE808IVE

MMCP Trade Twins — What We Built and What It Solves
The Problem
America is facing a silent crisis. Skilled trade knowledge is disappearing faster than it can be transferred.
300,000 electricians retire this decade. 200,000 plumbers. Ironworkers, welders, HVAC technicians, pipefitters — an entire generation of master craftspeople walking out the door, taking 20-40 years of hard-won expertise with them. No apprentice can absorb it fast enough. No robot has validated data to train on. The knowledge simply dies.
Current AI cannot solve this. Every AI agent today is stateless — it starts from zero on every call, has no identity, no accountability, no memory of what it built yesterday. You cannot preserve generational knowledge with an AI that forgets everything the moment the conversation ends.
What We Built
MMCP Trade Twins — a constitutional multi-agent orchestration system that agentically builds digital twin knowledge bases for every licensed skilled trade simultaneously, governed by one constitutional layer.
One command triggers 24 specialized agents in parallel — one per trade. Each agent is issued a unique identity, a blockchain wallet, and a set of constitutional rules it cannot violate. Nebius Token Factory runs Mixtral-8x7B inference on all agents simultaneously. Tavily pulls live authoritative trade codes and regulations in real time — not training data, live specs retrieved at the moment of knowledge formation. Every output is gated by Toloka human validators — engineering specialists who score whether the AI got it right. Only validated knowledge becomes a marble. Every marble is provenance-hashed, traceable to its source document, the agent that created it, and the human who approved it. Agent credit is written to the Scroll blockchain — good agents earn, bad agents are dissolved.
The result is a MAG — Memory Augmented Generation system. Not RAG. Each marble is multimodal, episodic, spatially indexed, and constitutionally governed. Journeyman knowledge becomes something you can navigate, query, and train robots on.
What It Solves
For humans: Apprentices learn from validated journeyman knowledge instead of hoping to find a master willing to teach. The great transfer of expertise — previously impossible at scale — becomes systematic infrastructure.
For robots: The Unitree G1 and every humanoid robot that follows doesn't train on internet scraping or synthetic data. It trains on constitutionally governed, human-validated, provenance-traced expert knowledge. That's a fundamentally different quality of training signal.
For the industry: Every licensed trade in America — 24 and growing — gets a living digital twin that builds itself through use, compounds institutional knowledge over time, and never loses what a retiring journeyman knew.
Why Nothing Else Does This
PermitFlow has 80% ops headcount because their AI can't handle edge cases. ChatGPT can answer a question about rack installation — it cannot build a governed knowledge base that trains a robot. LangChain orchestrates tasks. MMCP orchestrates agents — with identity, economics, governance, and memory baked into every call.
The constitutional layer is the breakthrough. Eight invariants that cannot be violated. No marble forms without human validation. No agent self-authorizes. No knowledge enters the training pipeline without a provenance chain traceable to the source. Robots don't learn from unverified data. Ever.

- **Team:** [Daniel Kaneshiro](https://cerebralvalley.ai/u/Cre808ive)
- **GitHub:** https://github.com/ChainMailGlobal/Nebius.Build
- **Demo video:** https://youtu.be/ND5YzeIidPc
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=18

### 49. ShadowDance

ShadowDance is a python package that enables observability, tracing, logging, and cost-tracking from the cloud to the claw with single line of code. Just because the robot drops the ball, doesn't mean you have to as well.

- **Team:** [Christopher McKenzie](https://cerebralvalley.ai/u/kristopolous)
- **GitHub:** https://github.com/kristopolous/ShadowDance
- **Demo video:** https://www.youtube.com/watch?v=4p-hL6cBj8Q
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=19

### 50. Framewise

Framewise cuts robot perception costs by 80-90%+ by never sending the same scene twice. Instead of flooding a VLM with redundant frames, it uses pixel-level and semantic change detection to gate inference - only the changed region gets sent, only when something actually moved. On change frames, a fast draft model (Gemma-27B) answers first; if uncertain, it escalates to Qwen-72B. The result: large-model accuracy where it matters, near-zero cost everywhere else.

It implements RAG, reranking, Tavily for more context for robot, streaming frames and measured speculative post draft routing based on certainty (confidence driven).

- **Team:** [Aishwarya Gune](https://cerebralvalley.ai/u/aishwaryagune)
- **GitHub:** https://github.com/agg111/framewise
- **Demo video:** https://youtu.be/XETQ567842Q
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=20

### 51. AgentCommerce

Every online purchase requires a human in the loop — browsing, comparing prices, checking if a seller is legit, negotiating. AgentCommerce
eliminates that entirely.

AgentCommerce is an autonomous commerce platform where two AI agents — a Buyer and a Merchant — negotiate purchases end-to-end with zero human
intervention. The Buyer agent researches real-time market prices, verifies the merchant's identity through digital credentials, and negotiates
the best deal below market price. The Merchant agent responds with counter-offers, price-matches competitors, and closes the sale — all in under
60 seconds.

What makes this different from anything else out there:

Heterogeneous multi-model negotiation — The Buyer runs on Nebius AI Studio (MiniMax-M2.1), the Merchant runs on OpenRouter (MiniMax-M2). Two
different AI providers negotiating a real transaction. This proves autonomous commerce works across infrastructure boundaries — agents don't need
to share the same platform to transact.

Verifiable digital identity — Before any money moves, both agents exchange and verify ISO 18013-5 mdoc credentials — the same standard behind
mobile driver's licenses and digital payment credentials (DPC). Trust is cryptographic, not assumed. This is built on real-world expertise from
the Open Mobile Hub initiative at the Linux Foundation.

Real-time price intelligence — Tavily Search pulls live pricing from official retailer sites (nike.com, garmin.com, rei.com). The buyer agent
uses verified market prices as a hard ceiling and negotiates 10-20% below. It never overpays.

Live visualization — Every step of the transaction streams via WebSocket to a React dashboard: Discovery, Research, Credential Exchange,
Negotiation, Agreement, Payment, Confirmed. Judges can watch two agents think, negotiate, and close a deal in real-time.

The result: a Garmin watch listed at $450, verified at $349.99 market price, autonomously purchased for $315 — 10% below market — with both
parties' identities cryptographically verified.

Today it buys running shoes. Tomorrow, warehouse robots procure their own replacement parts, fleet managers' AI agents negotiate fuel contracts,
and IoT devices purchase their own cloud compute. AgentCommerce is the protocol for an economy where machines transact autonomously and securely.
This is what commerce looks like when agents go beyond tool calls.

- **Team:** [Diego Zuluaga](https://cerebralvalley.ai/u/DiegoZ)
- **GitHub:** https://github.com/dzuluaga/agentcommerce
- **Demo video:** https://youtu.be/eeumOEpX1ik
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=26

### 52. Pool Watch

Drowning is the #1 killer of children aged 1–4 in America, and most of those deaths happen in backyard pools in the seconds a parent looks away. AI drowning detection exists, but it's built for hotels and aquatic centers — no family can afford it, and no one has built it for the home. Pool Watch is a privacy-first, in the future fully on-device detection pipeline that captures real-time biomechanical signs of drowning and alerts you instantly — no cloud, no subscription, no video ever leaving your property.

- **Team:** [Andrzej Firek](https://cerebralvalley.ai/u/andrzejarturr)
- **GitHub:** https://github.com/Muifrend/pool-watch
- **Demo video:** https://youtu.be/mTcqEFk1tbE
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=21

### 53. JC

Test if chaos engineering or provoking robots to use more chaotic exploration patterns to expand the potential behavior space has potential to speed up learning. Also see if natural language can be used to speed up training by assigning weights more intelligently that describe these exploration patterns: cautious, random, aggressive, strange.

- **Team:** [Janice Chan](https://cerebralvalley.ai/u/parallelmind)
- **GitHub:** https://github.com/tulang-nb2022/robotics-simulation-training-hackathon
- **Demo video:** https://youtu.be/oZdTVlnMGEM
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=22

### 54. SwitchCost

SwitchCost is an AI inference migration engine that helps teams safely move from expensive proprietary models (GPT-4, Claude, Gemini) to open-source alternatives on Nebius Token Factory while maintaining quality parity. It works in three steps: first it decomposes a complex agentic task into typed subtasks (classify, extract, summarize, reason, search, generate), then replays each subtask against all available open models across Nebius and OpenRouter while scoring quality via an LLM-as-judge, and finally recommends the cheapest model per subtask type that maintains >80% quality — projecting monthly savings at configurable call volumes. The result is a data-driven migration plan with a live dashboard showing a quality heatmap, cost comparison charts, and an interactive savings calculator, turning what used to be a risky gut-feel switch into a confident, evidence-backed decision.

- **Team:** [Ayush Ojha](https://cerebralvalley.ai/u/ayushojha)
- **GitHub:** https://github.com/Ayush10/switchcost
- **Demo video:** https://drive.google.com/drive/folders/1xDd1oYWUGaOGjoDDGtYCPLEL_bjQbrZT?usp=drive_link
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=25

### 55. Topology

10 teams of LLM agents compete on a task. Each team has a randomly generated organizational genome — hierarchy, communication pattern, decision-making style, work distribution, roles, and team size. The weakest teams dissolve. The strongest mutate and propagate. Over 5 generations, evolution finds the optimal agent architecture that no human would have designed.

- **Placement:** Finalist
- **Team:** [Nigel Robinson](https://cerebralvalley.ai/u/nigel_xavier)
- **GitHub:** https://github.com/nxrobins/topology
- **Demo video:** https://youtu.be/SIUD4TFAiNU
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=23

### 56. CHAI

CHAI (Companion Humanoid AI) is a real-time companion system for the Unitree G1 humanoid robot that detects approaching humans, greets them and autonomously clears obstacles from the path using the robot's arm, acting as a guide for visually impaired users. It uses Nebius AI Studio's VLM (Qwen2.5-VL-72B) for vision and LLM (Llama-3.3-70B) for decision-making, and runs in both MuJoCo simulation.

- **Team:** [Vatsal Bajaj](https://cerebralvalley.ai/u/vb99), [Jake Kang](https://cerebralvalley.ai/u/kangjake)
- **GitHub:** https://github.com/galileo-ml/nebius-hack
- **Demo video:** https://vimeo.com/1173842163?share=copy&fl=sv&fe=ci
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=24

### 57. Team Big Floofy Cat

Doctor Auditor is a local-first AI system for reviewing sensitive conversations without centralizing raw recordings or transcripts. Most audio AI workflows start by uploading the entire session to the cloud, which creates unnecessary privacy and control risks before anyone has approved what leaves the device. Doctor Auditor reverses that pattern: raw audio, full transcripts, draft findings, and review state stay on the local machine, and only approved artifacts are allowed to cross the boundary.

- **Team:** [Robert Cordwell](https://cerebralvalley.ai/u/discordwell)
- **GitHub:** https://github.com/discordwell/doctor-auditor
- **Demo video:** https://www.loom.com/share/8a7424fcce234c60b5765bca66de93c4
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=27

### 58. arm-pilot

Autonomous Task Discovery System

When a robot receives a job command, a Nebius-powered LLM inference engine autonomously searches the internet to identify the most suitable task for that job. The corresponding motion data is then retrieved online and deployed directly to the robot.

In our current demo, the humanoid arm executes pick-and-press operations in real time.

Going forward, robots will autonomously discover and acquire the necessary tasks and motions from the internet upon receiving a single command — enabling full job completion across any worksite without manual programming. used tavily for grounding search.

- **Team:** [Sota Miyajima](https://cerebralvalley.ai/u/Sota), [Munenori Saito](https://cerebralvalley.ai/u/mune), [Kouta Ueda](https://cerebralvalley.ai/u/Kota), [Kush Ise](https://cerebralvalley.ai/u/KUSH2704)
- **GitHub:** https://github.com/Sota-sota/nebius-build-sf-2026/tree/kush
- **Demo video:** https://www.loom.com/share/1b3b7a70740c46a08b600e5d7c622a4b
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=28

### 59. Asha

Asha is a WhatsApp-native AI assistant that helps small home-based food businesses manage orders and remember customers — all through WhatsApp, with no apps or dashboards to learn. Small home food businesses (meal prep services, home bakers, pop-up kitchens) in the US rely on WhatsApp groups, Instagram DMs, and manual tracking, leading to missed orders, forgotten special requests, and growth capped by what owners can mentally track. Existing solutions liike DoorDash cost $30–250/month and target large brands, pricing out micro-businesses. Ash is an AI assistant that lives in WhatsApp — business owners chat with Ash to set up their business, and customers place orders by texting Ash directly. Ash remembers customer preferences, sends proactive reminders, and provides daily business summaries, turning WhatsApp into a complete order management system.

- **Placement:** Finalist
- **Team:** [Geoffrey Ju](https://cerebralvalley.ai/u/xmens), [ranya belmaachi](https://cerebralvalley.ai/u/ranyabel), [Mohammed Misran](https://cerebralvalley.ai/u/misran)
- **GitHub:** https://github.com/misran3/asha
- **Demo video:** https://drive.google.com/file/d/1RjiQio_4Iy6TRSmWpwu5q8mLSmsm8VXk/view?usp=sharing
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=34

### 60. AutoRobots

An autonomous AI scientist that teaches simulated robots to move. No human reward engineering. The LLM agent iterates on reward functions, trains RL policies, evaluates results, and accumulates improvements — all on Nebius infrastructure.

- **Team:** [Warner Wu](https://cerebralvalley.ai/u/warner), [ASTRAL LABS](https://cerebralvalley.ai/u/Astral)
- **GitHub:** https://github.com/anjan04/nebiushack
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=29

### 61. ShopAgent

We hated when I see people wearing same outputs in SF. 

ShopAgent is an AI-powered shopping assistant to level the playing field between big retailers and small businesses. We believe independent makers and local shops deserve the same visibility as Amazon. ShopAgent searches both worlds simultaneously, ranks every result by price, shipping speed, quality, and ethical sourcing, and presents them side by side.

ShopAgent makes it just as easy to discover and support a small business as it is to click "Buy Now" on Amazon.

- **Team:** [Victor Park](https://cerebralvalley.ai/u/victorjhp), [Lucas Georgescu](https://cerebralvalley.ai/u/lucasg), [Diego Viladoms](https://cerebralvalley.ai/u/Diego_V), [Shijie Xu](https://cerebralvalley.ai/u/shijiexu)
- **GitHub:** https://github.com/jihwan-victor-park/NebiusHack
- **Demo video:** https://youtu.be/aABH9wlcX1s
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=30

### 62. Injester

Injester is an AI-native engine that transforms "agent-hostile" websites into structured data interfaces. While the web was built for human attention, our tool bridges the gap for LLMs by stripping 90% of token noise using Tavily and restructuring content via Nebius. Featuring a "Karpathy Loop," the system autonomously iterates to improve task accuracy from $2/5$ to $5/5$. It’s the essential layer for Agent Experience Optimization (AEO).

Link to slides:
http://injester.com/injester.pdf

- **Placement:** 1st Place
- **Team:** [Alex Shirazi](https://cerebralvalley.ai/u/phlint), [Vishal Verma](https://cerebralvalley.ai/u/slowpoison), [Benjamin Shyong](https://cerebralvalley.ai/u/BenjaminBear)
- **GitHub:** https://github.com/InjesterLol/Main
- **Demo video:** https://youtu.be/ZRyhK3GvtV0
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=33

### 63. AutoEvolve

Nebius Token Factory powers the cloud-side "task planner" by running Qwen 2.5 as a reasoning layer that takes in visual and natural language input — such as "pick up the cup" — and decomposes it into a series of fine-grained manipulation steps for robot arms, like approaching the object, opening the gripper, grasping, and lifting.

A fine-tuned GR00T N1.6-3B consumes each step to produce continuous joint actions via its Vision Language Model (VLM), enabling the robot to perceive its environment, reason about what's happening, and execute meaningful physical actions across tasks like reach-and-grasp, object handoff, and basic manipulation.

- **Team:** [Miaosen Zhou](https://cerebralvalley.ai/u/then), [Lily Su](https://cerebralvalley.ai/u/lilyxsu), [Long Haitao](https://cerebralvalley.ai/u/takeiteasydragon)
- **GitHub:** https://github.com/zhoumiaosen/AutoEvolve
- **Demo video:** https://youtu.be/YQ5lV0-kl68
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=31

### 64. voiceClaw

One-liner
Connect open-source voice hardware (Xiaozhi ESP32) to OpenClaw Agent Gateway — plug in, speak, get AI Agent response — with lightweight server implementations in Python, Go, and Swift.
Problem
AI Agents today are text-only (web, app, API). Real-world scenarios need voice: cooking in the kitchen, working on a factory floor, elderly users who can't type. Existing voice assistants (Alexa, Siri) are closed ecosystems — developers can't plug in their own agents.
Solution
We bridge OpenClaw Agent Gateway with Xiaozhi ESP32 open-source voice hardware using an LLM Passthrough architecture:

- **Team:** [zhuang hua](https://cerebralvalley.ai/u/aadebuger)
- **GitHub:** https://github.com/aadebuger/voiceclaw-go
- **Demo video:** https://youtube.com/shorts/0fqLpK6F_mI
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=32

### 65. hideNseek

Labs are training and have found performance boost training llms on games, we want to do the same with robots in a classical childhood game of Hide and Seek!

Robots: Seeker and Hider use vision models to Rl policy to train robots to interact dynamically with their environment and change hiding strategies based on live feedback and interaction.

- **Team:** [aTG R](https://cerebralvalley.ai/u/atg)
- **GitHub:** https://github.com/r-agni/hideNseek
- **Demo video:** https://drive.google.com/drive/folders/1EE_EHp2BXPUbsmDWurxY6EKqiKHKRU3F?usp=sharing
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=35

### 66. rOCDbot

We use llm familiarity with OCD to make higher precision robots. The robot sees the object is out of place, does actions in a loop to get the object to be in a perfect location, this trace of original position and actions are used in RL training

- **Team:** [Kirill Igumenshchev](https://cerebralvalley.ai/u/kirilligum)
- **GitHub:** https://github.com/kirilligum/rOCDbot-cerebral-valley-hackathon-260315
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=49

### 67. Movebot

Move!

- **Team:** [Carlos Lara](https://cerebralvalley.ai/u/carloslara)
- **GitHub:** https://github.com/LaraArias/MoveBot
- **Demo video:** https://www.loom.com/share/91d5e377750d4936af4f99affc138a76
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=52

### 68. Kleiner Bot

Teach robots emotion (by reading my own)

Edit: I'm withdrawing this project now. Didn't have enough time to record.

- **Team:** [Tor](https://cerebralvalley.ai/u/tor)
- **GitHub:** https://github.com/hagemt/reachy-emotes
- **Demo video:** https://linkedin.com/in/hagemt
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=66

### 69. zippling

Agentic E-commerce customer service resolution and management platform that automatically handles customer support resolution agentically based on platform policies along with text and visio

- **Team:** [Parth Gala](https://cerebralvalley.ai/u/parthgala)
- **GitHub:** https://github.com/ParthGala2k/zippling_ai_commerce
- **Demo video:** https://youtu.be/8Kn2S8PlUwo
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=71

### 70. Robots For Social Good

Robotics for Social Good is a multimodal humanoid robotics system that learns socially beneficial behaviors from real-world human demonstrations. The platform allows users to upload videos of positive actions such as helping others, cleaning environments, or assisting people in daily tasks. These videos are transformed into structured training data that teaches a humanoid robot how to perceive its surroundings, understand intent, reason about possible outcomes, and execute meaningful physical actions.

At the core of the system is a multimodal perception pipeline that processes visual, audio, and contextual information from uploaded videos. The data is annotated through Nomadic ML, which enables scalable labeling of objects, human poses, interaction events, and social intent. These annotations are converted into structured datasets containing scene states, action trajectories, and reward signals, forming the foundation for training the robot’s decision-making models.

The system uses a Vision-Language-Action (VLA) architecture that connects perception with physical control. Visual observations are encoded through transformer-based vision models and combined with language grounding to interpret the meaning of actions within a scene. This fused representation allows the system to generate action policies that translate environmental understanding directly into motor commands for a humanoid robot.

To ensure reliable real-world behavior, the robot policies are trained using reinforcement learning. The training process optimizes for task success, safety, stability, and social usefulness. The reward structure encourages the robot to perform actions that align with positive human outcomes, such as assisting a person or correctly handing over an object. Tasks used during training include navigation, obstacle avoidance, reach-and-grasp, object delivery, and cooperative manipulation.

A predictive world model is integrated into the system to simulate future states of the environment before actions are executed. By forecasting potential outcomes, the robot can evaluate multiple strategies and choose the action that maximizes expected success while maintaining safety and efficiency. This allows the robot to reason about complex scenarios instead of relying solely on reactive behavior.

Top-performing policies can be deployed on humanoid platforms such as the Unitree G1, enabling real-world demonstrations of socially beneficial robotic behavior. The system continuously improves through a feedback loop where robot execution data is added back into the training pipeline, refining perception models, action policies, and decision-making strategies over time.

An additional feature of the platform is its community contribution mechanism, where users who upload impactful demonstrations receive points when their examples help improve robot performance. This transforms crowdsourced human kindness into machine-learnable knowledge, creating a collaborative ecosystem for training robots that benefit society.

By combining human demonstrations, multimodal perception, reinforcement learning, and humanoid robotics, Robotics for Social Good introduces a scalable framework for developing intelligent robots capable of assisting people and contributing positively to real-world environments.

- **Team:** [Deon Menezes](https://cerebralvalley.ai/u/Deonmenezes)
- **GitHub:** https://github.com/deonmenezes/robotics-for-social-good
- **Demo video:** https://youtu.be/_yBrWxeqdzc
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=36

### 71. Should-work

Should-work makes sure your claude code work day and night.

It  is an AI agent orchestration platform for solo developers. It manages a fleet of Claude Code agents to build software, so that the human prioritizes issues and reviews PRs, agents do the implementation.

- **Team:** [Julaiti Alafate](https://cerebralvalley.ai/u/julaiti)
- **GitHub:** https://github.com/arapat/should-work
- **Demo video:** https://youtu.be/L4Osibyef6M
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=39

### 72. Sous Bot

An assistive grocery robot for visually impaired and elderly users. It helps in picking items from a grocery store based on the meals you want

- **Team:** [Sathvick Reddy Narahari](https://cerebralvalley.ai/u/SathvickN), [divya Narahari](https://cerebralvalley.ai/u/divyanarahari97), [Jyothirmai Kottu](https://cerebralvalley.ai/u/jkottu)
- **GitHub:** https://github.com/DivyaNarahari97/sous-bot/tree/main
- **Demo video:** https://drive.google.com/drive/folders/1M9op3kYpl50Lv3DrdB-zrQle3t7O0HLP?usp=drive_link
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=48

### 73. AI Warehouse Fulfillment

An LLM-powered warehouse robot that reasons, routes, and adapts to chaos — in real-time.

What happens when you give a reasoning LLM control of a warehouse fulfillment station? It analyzes incoming packages, plans optimal routing, executes step by step — and when rush orders flood in, it re-plans on the fly.

- **Team:** [Nelson Lai](https://cerebralvalley.ai/u/chineseman)
- **GitHub:** https://github.com/chinesepowered/hack-neb
- **Demo video:** https://www.youtube.com/watch?v=p85FhbB4mPA
- **Project:** https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery?project=37

---

Markdown version of https://cerebralvalley.ai/e/nebius-build-sf/hackathon/gallery. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt