precision_manufacturing Sideline
Player 1
0
0 games · 0 sets
0 – 0
Games
Ready
Player 2
0
0 games · 0 sets
FRAME 000 · READY
Frame 0
smart_toy Sideline
idle
psychology
Agent Console
0 events

How Sideline Works

An agentic AI referee that watches sports video, reasons about every play using chain-of-thought, and acts through any physical form — from dashboards to robots.

visibility

Observe

Extract frames from live video or recordings at configurable FPS. Each frame becomes input for the vision model.

psychology

Perceive

Nebius Token Factory VLM analyzes each frame with game context. Qwen2-VL-72B or Cosmos Reason 2.

balance

Reason

Chain-of-thought reasoning: what happened → which rule applies → what's the correct call → confidence level.

target

Decide

The model selects tools to call: update_score, announce_call, robot_gesture, or no_call.

smart_toy

Act

Execute actions through three protocols: Function Calls, MCP servers, and A2A agent communication.

memory

Remember

Maintain game state across the entire match. Score, rally context, call history — the agent never forgets.

Agent Pipeline

Video Feed ──▶ Frame Extraction (1 fps)
                    │
                    ▼
            Nebius VLM + Tool Definitions
            (vision + game state context)
                    │
                    ▼
            ┌── Tool Calls ──────────────────────┐
            │                                     │
     ┌──────┴──────┐  ┌──────────┐  ┌───────────┴───────┐
     │ Function     │  │   MCP    │  │       A2A         │
     │ Calls        │  │  Server  │  │  Agent-to-Agent   │
     │ (internal)   │  │(external)│  │  (multi-agent)    │
     └──────┬──────┘  └────┬─────┘  └───────────┬───────┘
            │              │                     │
            ▼              ▼                     ▼
     ┌─────────────────────────────────────────────────┐
     │  Score Update · Voice · Dashboard · Robot · API  │
     └─────────────────────────────────────────────────┘

Three Action Protocols

Function Calls

Internal Tools

The VLM model decides which tools to invoke. Score updates, announcements, and robot gestures — all through native tool calling.

MCP

Model Context Protocol

Expose referee capabilities as standardized MCP tools. Any MCP client — Claude, ChatGPT, Cursor — can use Sideline as a referee tool.

A2A

Agent-to-Agent Protocol

Multiple Sideline agents covering different cameras coordinate decisions via Google's A2A protocol. Like VAR, but autonomous.

Physical Embodiment

Dashboard

Video + reasoning + scoreboard. Always works, zero hardware needed.

MentorPi

Tank bot with camera, speaker, lidar. Powered by NASA JPL's ROSA framework.

SO-101 Arm

6-DOF robotic arm via HuggingFace LeRobot. Signals calls with physical gestures.

Unitree G1

Humanoid referee. Walks the court, gestures, speaks. The end goal.

Team Sideline

Built at Nebius.Build SF — March 15, 2026

RJ

Ravinder Jilkapally

Agent Pipeline + Dashboard

Edge AI, full-stack, inference optimization. Built the end-to-end agentic pipeline.

LinkedIn →
VR

Vivek Gopal Ramaswamy

Robotics + Architecture

CMU Robotics, AI PM. Self-driving systems, VLA models, embodied AI hackathon mentor.

LinkedIn →
KH

Kruthik Hulisandra

Domain Rules + Agent Logic

Actual sports referee. Managed 14 leagues, 6000 participants. AI engineer with MCP expertise.

LinkedIn →
VB

Visshwa Balasubramanian

Backend Engine + Model Work

Palantir, USACO Gold, NeurIPS workshop paper. CS heavy hitter.

LinkedIn →

Tech Stack

Inference
Nebius Token Factory
Backend
Python + FastAPI
Frontend
Vanilla JS + WebSocket
Protocols
Tool Calling + MCP + A2A
Robot
MentorPi + ROSA (NASA JPL)
Models
Qwen2-VL-72B / Cosmos R2