Lack of standardized AI in regional events
More than 130,000 regional MMA events run annually in the US with no standardized AI scoring or computer-vision analytics layer.
Prepared for NVIDIA Inception review
Thank you for judging, Fight Judge AI.
This page is built for NVIDIA Inception reviewers. It starts with the market gap, shows what is already live, then isolates the GPU bottlenecks, revenue path, and long-term compute trajectory behind Fight Judge AI.
Fight Judge AI is building the AI command layer for live combat sports: real-time scoring support, post-event auditability, fighter intelligence, and computer-vision infrastructure for an industry that still runs with minimal standardized data systems.
The problem
An underserved industry is entering its data era.
Outdated judging in major leagues
The biggest organizations in combat sports still rely on subjective human judging systems that were not designed for modern real-time evidence workflows.
Limited existing CV solutions
Existing vendors are narrow: boxing-only analytics, limited grappling coverage, or partial tools without a complete operational platform.
Fighters lack structured performance data
Training, gameplanning, and opponent preparation remain largely manual because reliable fight telemetry is still difficult to capture and normalize.
Promotions and matchmakers operate blind
Regional promotions still make talent, matchmaking, and event decisions without consistent intelligence across fighter history, style, and performance trends.
Sportsbooks have minimal real-time fight data
Combat sports is a fast-growing betting category, but the data infrastructure needed for live APIs and structured feeds is still early.
The platform
The operating system layer already exists.
ScoreOps
Real-time AI-assisted fight scoring and live stats for promotions and event operators.
FightAudit
Post-event judge transparency, scoring comparison, and timestamped evidence workflows.
FighterIQ
Performance scoring, scouting reports, and matchmaking intelligence built on structured fight data.
FightVault
Searchable archive of fighters, fights, events, and derived metrics that becomes more defensible with every show.
FanJudge
Live fan scoring and event engagement surfaces that create additional promotion value on fight night.
Fight Mentor AI
AI-powered fighter development, clip-grounded technique analysis, predictive matchmaking, and AI-generated game plans.
Next big thing in development
Fight Mentor AI is the agentic layer that turns this data moat into daily fighter workflows.
Multi-step fight reasoning, not a single-model chatbot
Fight Mentor AI is being developed as an agentic system that can break opponent prep, training review, and gameplan generation into chained tasks with grounded outputs instead of one-shot responses.
Best-model routing by task
We are designing the next layer to route across multiple LLMs so video understanding, retrieval, coaching synthesis, and structured report generation each use the model class best suited to the job.
Fight-specific retrieval with evidence
The retrieval layer is being built around fight clips, round notes, opponent tendencies, event history, and structured scoring data so every coaching answer can be tied back to actual evidence.
Pull likely patterns from prior fights, summarize openings, and generate a clip-backed scouting report for coaches and fighters.
Convert retrieval plus fight-style analysis into draft tactical plans by phase, position, and round context.
Track sparring notes, conditioning logs, weight-cut checkpoints, and coach instructions in a fighter-specific memory layer.
Let fighters ask questions against their own rounds and receive answers grounded in exact moments, not generic advice.
Leveraging NVIDIA technology
GPU acceleration is the direct unlock for scale.
Fight Judge AI is already processing post-event fight footage with YOLOv8-pose, extracting pose keypoints and structured signals for scoring and analysis. The next phase is not speculative product design. It is replacing CPU bottlenecks with a real NVIDIA acceleration stack.
Processing 1,500+ fight videos for strikes, grappling states, knockdowns, and pose extraction into structured JSON for scoring and review workflows.
Video frame extraction and pose inference, combat model fine-tuning, broadcast inference, fighter database analytics, and technique CV all improve materially with GPU access.
These are the tools that map directly to Fight Judge AI's next deployment phases, not optional future experiments.
Move the current post-event CV workflow from CPU drag into repeatable GPU throughput.
Train on large-scale annotated combat frames instead of limiting iteration speed to infrastructure gaps.
Use DeepStream for live event and broadcast-grade CV pipelines once the base inference stack is accelerated.
RAPIDS-backed analytics can accelerate database exploration, report generation, and historical analysis.
TensorRT matters for both venue-side deployment and persistent production efficiency.
T4 edge plus DeepStream opens the path to phone-camera and gym-floor technique feedback.
Traction
Live, deployed, and already compounding.
- Deployed with Peak Fighting Championship across three live events.
- ScoreOps already used in kiosk mode at sanctioned events.
- Post-event pipeline runs from scoring to review to archive layers.
- Historic milestone: an AI-assisted live scoring system already deployed in sanctioned US MMA events.
- 1,500+ full-fight MP4 videos.
- 13.5k+ fight logs.
- 1,250+ events.
- 2,600+ scored fighters with pre and post-fight intelligence reports.
- An actively labeled combat-sports CV dataset.
- Anchor client in production today.
- $15,000 generated across the first three events.
- Early event model operating above 60% net margin.
- Expansion path into larger promotions, broadcasters, and fight-data buyers.
GPU compute roadmap
A phased path from backlog processing to persistent multi-event inference.
Phase 1: NVIDIA T4
Accelerate YOLOv8-pose post-event inference across the current fight backlog and ongoing event intake.
Phase 2: NVIDIA A100 (40-80GB)
Fine-tune a combat-specific pose model on roughly three million annotated frames. Inception GPU access is critical here.
Phase 3: Multi-A100
Expand into strike classification, grappling recognition, and fighter technique CV across 13k+ fights.
Phase 4: Persistent H100
Support real-time DeepStream inference for simultaneous live events and phone-camera technique feedback.
Multi-GPU + Blackwell GB200
Persistent multi-event inference workloads for live scoring, broadcast overlays, training intelligence, and data APIs.
Market and business model
Three underserved markets, one compounding infrastructure layer.
- $2-3B estimated addressable market by 2028 across promotion analytics, fighter intelligence SaaS, and fight data APIs.
- 130,000+ US regional events as the beachhead where data infrastructure is still minimal.
- 5M+ US athletes and 30M+ global athletes in the fighter and gym expansion market.
- Sports betting as the upside layer once reliable real-time fight data becomes available.
- Promotion platform packages from $1,500 to $4,500 per month, plus per-event SKUs.
- Fighter analytics and training app positioned at $29 to $99 per month.
- Gym SaaS positioned at $99 to $349 per month.
- FightVault API, opponent intelligence reports, and matchmaking licensing as higher-scale data monetization layers.
- Current cost per event is approximately $1.25 in GPU and hosting.
- Revenue per current promotion contract ranges from $1,500 to $4,500 monthly plus add-ons.
- Gross margin expands as GPU cost is amortized across more events, models, and downstream products.
Founder fit
Built by an operator with both domain and technical depth.
Eric Gann is a former professional MMA fighter with 18 years of combat-sports experience and the full-stack builder behind Fight Judge AI. The platform, backend, CV pipeline, and live scoring workflows were all built with direct knowledge of judging, fight dynamics, and fighter needs.
- Python
- FastAPI
- React 19
- PostgreSQL
- YOLOv8
- n8n
- WebSocket
- Azure Blob Storage
Why NVIDIA Inception
Requesting support for a live product, not a hypothetical build.
- A combat-sports CV and AI platform category that is still largely open.
- A proprietary labeled fight-video and structured event dataset.
- Live production deployment with real revenue already generated.
- A direct path to promotions, broadcasters, sportsbooks, and data customers.
- A100 access for combat model fine-tuning.
- TensorRT optimization support and technical training.
- DeepStream guidance for live broadcast CV.
- RAPIDS support for fighter database analytics.
- Preferred GPU access paths as compute demand compounds.
- Ecosystem introductions relevant to sports-tech growth.
Commitment
Fight Judge AI is designed to become a long-term GPU customer.
As the platform scales from post-event analysis into live multi-event inference, fighter technique CV, and data APIs, compute demand compounds. Inception is the right starting point because the technical relationship matters as much as the compute itself.