Multi-agent reasoning · Consensus scoring · Autonomous execution
NemoFish is a swarm intelligence platform that coordinates multiple AI agents to generate probabilistic forecasts for competitive sports events. Each agent brings a different analytical lens — statistical modeling, psychological profiling, market signal analysis, scenario simulation — and they reach consensus through weighted voting. The system then scores opportunities by expected value and can execute positions on prediction markets autonomously.
This project began as a fork of MiroFish, an open-source multi-agent simulation engine developed by the CAMEL-AI research community. We extended MiroFish with a full-stack sports analytics pipeline and a unified control room for real-time monitoring.
An R&D project by ASG Compute — exploring whether coordinated multi-agent reasoning can identify inefficiencies in prediction markets.
| Metric | Value |
|---|---|
| Historical matches analyzed | 971,320 |
| Unique players profiled | 33,962 |
| Tournaments covered | 9,201 |
| Years of ATP data | 1968 – 2026 (57 seasons) |
| Surfaces modeled | Hard · Clay · Grass · Carpet |
| Live data feeds | 4 providers (api-tennis, Sportradar, The Odds API, Polymarket) |
| AI agents in swarm | 6 specialized reasoning agents |
| Features per match | 47 (ELO, surface form, H2H, momentum, odds, news) |
| Codebase | 30,000+ lines across 160 modules |
The system ingests nearly 1 million historical matches to build surface-weighted ELO ratings and train XGBoost probability models via walk-forward validation. At runtime, the 6-agent swarm overlays real-time signals (live odds movement, injury news, scheduling fatigue) on top of the statistical baseline to generate final consensus predictions.
| Layer | Description | Key Components |
|---|---|---|
| Data Ingestion | Real-time feeds from 4+ providers | api-tennis.com, Sportradar, The Odds API, Polymarket |
| Quantitative Models | Surface-weighted ELO, XGBoost probability engine | Walk-forward validation, head-to-head analysis |
| Agent Swarm | 6 specialized reasoning agents | Statistical, Psychological, Market, Contrarian, News, MiroFish (LLM) |
| Consensus Engine | Weighted vote aggregation with edge scoring | Configurable thresholds, data quality gates |
| Execution Pipeline | Kelly criterion position sizing, risk management | Polymarket CLOB integration, PnL tracking |
| Control Room | Unified real-time dashboard | Live scores, odds movement, tomorrow's slate, trade journal |
 |
| KPIs, live event feed, odds movement tracking, agent consensus heatmap |
 |
| Tomorrow's slate — win probabilities, edge calculations, and data quality scores per match |
DATA FEEDS INTELLIGENCE OUTPUT
────────── ──────────── ──────
api-tennis.com ──┐
sportradar ──┤ ┌──────────────────────┐ ┌─────────────┐
the-odds-api ──┼──▶│ 6-AGENT SWARM │──▶│ CONSENSUS │
polymarket ──┤ │ │ │ ENGINE │
news feeds ──┘ │ ELO · XGBoost │ │ │
│ Psych · Market │ │ Edge ≥ 3%? │
│ Contrarian · LLM │ │ DQ check? │
└──────────────────────┘ └──────┬──────┘
│
▼
┌─────────────┐
│ EXECUTION │
│ │
│ Kelly sizing │
│ Risk limits │
│ CLOB orders │
└──────┬──────┘
│
▼
┌─────────────┐
│ DASHBOARD │
│ │
│ Live scores │
│ Odds tracker │
│ Trade journal│
└─────────────┘
Pipeline flow: Every cycle, the system ingests live data → builds a match dossier → runs each agent independently → aggregates into a consensus score → filters by minimum edge and data quality → sizes the position via fractional Kelly criterion → executes on prediction markets → tracks performance.
Six agents analyze each match independently before voting:
| Agent | Analytical Focus | Data Sources |
|---|---|---|
| 📊 Statistical | Surface-weighted ELO, recent form, head-to-head record | Sackmann historical data, api-tennis |
| 🧠 Psychological | Momentum, pressure response, fatigue patterns | Recent match sequences, scheduling |
| 💹 Market | Line movement, sharp money indicators | The Odds API, Polymarket orderbook |
| 🔄 Contrarian | Public bias detection, mean reversion | Reverse market signals |
| 📰 News | Injury reports, fitness status, travel burden | RapidAPI news feed |
| 🐡 MiroFish | LLM scenario simulation, qualitative reasoning | DeepSeek V3.2 via NVIDIA NIM |
Agents vote independently. The consensus engine applies calibrated weights and only surfaces opportunities above configurable thresholds (default: ≥3% edge, ≥35% data quality).
| Tool | Version | Check |
|---|---|---|
| Node.js | 18+ | node -v |
| Python | ≥3.11 | python --version |
| uv | latest | uv --version |
git clone https://github.com/ASGCompute/NemoFish.git
cd NemoFish
cp .env.example .env
# Edit .env — add your API keysnpm run setup:allnpm run dev
# Frontend → http://localhost:3000
# Backend → http://localhost:5001# API server
cd terminal && python api/dashboard_server.py &
# Dashboard UI
cd terminal/dashboard && npm run dev
# Open http://localhost:5173
# Generate predictions for tomorrow's matches
python intelligence/slate_runner.py --dry-runSee .env.example for the full list of required API keys.
NemoFish/
├── frontend/ # MiroFish simulation UI (Vue 3)
├── backend/ # MiroFish simulation engine (Python/CAMEL-AI)
├── terminal/ # Sports prediction terminal (our extension)
│ ├── agents/ # 6-agent swarm logic
│ ├── api/ # Dashboard REST API
│ ├── dashboard/ # Unified control room (React + Vite)
│ ├── execution/ # Position sizing & market execution
│ ├── feeds/ # Live data ingestion
│ ├── intelligence/ # Slate runner, scenario engine, dossier builder
│ ├── models/ # ELO engine, XGBoost classifiers
│ └── strategies/ # Kelly criterion, value confirmation
├── .env.example # Environment variable template
└── docker-compose.yml # Docker deployment
System parameters are defined in terminal/config.yaml:
bankroll:
initial_usd: 20
kelly_fraction: 0.25 # Quarter Kelly (conservative)
min_edge_percent: 3.0 # Minimum edge to act
max_bet_percent: 5.0 # Maximum position size
swarm:
tennis_agents: 6
consensus_threshold: 0.08 # Minimum edge from agent consensus
execution:
paper_trading: true # Paper mode by defaultNemoFish is built on top of MiroFish by @666ghj, backed by Shanda Group. The core simulation engine is powered by OASIS from the CAMEL-AI research community.
We extended MiroFish with a quantitative sports analytics pipeline, a multi-agent prediction swarm, and a live execution layer with a unified control room. All credit for the original simulation framework goes to the MiroFish team.
GPL-3.0 — consistent with the upstream MiroFish project. See LICENSE.
An ASG Compute Research Project
If you find this useful, give it a ⭐