Nebius.Build SF 2026

An AI referee.
Watch it
make the call.

240 million tennis matches played every year. Almost none have a referee. Sideline is a working prototype — an agent that watches video, makes calls with chain-of-thought reasoning, and signals a physical robot.

Prototype · Nebius.Build SF

SO-101 Arm — Ready

240M+

Matches Annually

Tennis alone. Add pickleball, volleyball, cricket — billions of plays with zero officiating.

~0%

Have a Referee

Amateur sports are completely unserved. Every close call is an argument.

3

Action Protocols

Function calls, MCP, and A2A — the agent acts through any interface, any robot, any client.

How It Works

Five stages. One agent loop. Video in, robot gesture out.

videocam

Video

Frame-by-frame capture from any camera source

visibility

VLM

Nebius Qwen2-VL-72B reads the frame, describes the play

psychology

CoT

Chain-of-thought: rules applied, confidence scored, call decided

gavel

Call

Score updated, voice announcement triggered, scoreboard synced

precision_manufacturing

Robot

Gesture executed — point out, raise arm, wave off — on the physical body

Full Pipeline

Video → Nebius VLM → CoT Reasoning → Tool Call → Score + Voice + Robot Gesture

Observe → Perceive → Reason → Decide → Act → Remember

No training data. No fine-tuning.

The VLM already understands what it sees. We give it the rules of the sport as context and let it reason.

0

Training Videos Needed

Qwen2-VL-72B on Nebius already understands visual scenes. We inject the sport rules as a system prompt. That's it.

6h

From blank repo to working demo

Fresh repo at 10:40 AM. Agent reasoning on live video by 5:00 PM. Domain knowledge in our heads, every line of code written at the event.

3

Action Protocols

Function calls for internal actions, MCP so any AI client can use our referee as a tool, A2A for multi-agent coordination.

The Agent Loop

Observe → Perceive → Reason → Decide → Act → Remember

Each frame feeds the loop. Game state persists across the match. The agent remembers what happened three points ago.

Sport Modules

Tennis is working. The rules engine and prompts are designed to swap — one file per sport. Pickleball and volleyball are next.

sports_tennis

Tennis & Table Tennis

Fault, out, ace, winner, let — full scoring with deuce and advantage. Ball tracking and line call analysis.

gavel

Built at hackathon

Fault, out, ace, let — full scoring with deuce and advantage.

smart_toy

Robot gestures

Point out, raise for ace, wave off for let — on physical hardware.

psychology

Chain-of-thought

Every call is reasoned and auditable. See the thinking in the dashboard.

sports_handball

Pickleball

Kitchen violations, line calls, scoring — the fastest-growing sport in America and still no referee at most courts.

schedule

Coming next

Same agent, same architecture. Swap the rules file and the sport changes.

sports_volleyball

Volleyball

Line calls, rotation tracking, serve analysis — the agent doesn't care what sport it's watching.

schedule

On the roadmap

Rules and prompts are pluggable. Tennis proved the pattern.

Same Brain. Any Body.

The agent doesn't care what it's running on. Swap the robot backend with one environment variable.

directions_car

Tier 1

MentorPi

Live

HiWonder tracked rover with mecanum wheels, camera module, speaker, and ROS2. Mobile, affordable, ready.

  • · Mecanum wheels — moves in any direction
  • · Built-in speaker for TTS announcements
  • · Lidar + depth camera
  • · ROS2 + Python control stack
Confirmed at event
precision_manufacturing

Tier 2

SO-101 Arm

LeRobot / HuggingFace 6-DOF robotic arm. Signals calls with precise arm gestures — point out, raise for ace, wave off for let.

  • · 6x Feetech STS3215 servos
  • · Leader-follower teleoperation
  • · LeRobot SDK — pip install lerobot
  • · Fault / out / ace / winner gestures
accessibility_new

Tier 3

Unitree G1

Fleet

Full humanoid. Walks to the line, raises an arm, makes the call. The closest thing to a real referee on the court.

  • · Full-body walking and gesturing
  • · Head tracking for situational awareness
  • · Unitree SDK — shared fleet at event
  • · A2A protocol for multi-agent coordination

Built With

Nebius Token Factory Qwen2-VL-72B MCP A2A Protocol FastAPI + WebSocket LeRobot ROS2 Three.js OpenCV

Why an Agent, Not Just Vision?

An image classifier can tell you the ball landed out. An agent knows the score, the server, the rules, and what happened three points ago.

Traditional CV

  • Ball in/out — that's it
  • No understanding of game rules
  • Can't track score or match state
  • Needs thousands of labeled training images
  • One sport per model

Sideline Agent

  • Understands rules, context, and match history
  • Tracks score across an entire match
  • Shows reasoning — auditable, challengeable
  • Zero training — uses pre-trained VLM + prompts
  • Swap sport by changing the rules file

The Team

Built at Nebius.Build SF 2026

6 hours. SHACK15, San Francisco.

RJ

Ravinder Jilkapally

Demo UI · Agent pipeline

LinkedIn →
VR

Vivek Gopal Ramaswamy

Robotics · Architecture · Pitch

LinkedIn →
KH

Kruthik Hulisandra

Domain rules · Agent logic

LinkedIn →
VB

Visshwa Balasubramanian

Backend engine · Robot control

LinkedIn →

Watch it make the call.

Tennis match. Reasoning agent. Physical robot. Built in 6 hours, open source — try it or read the code.