TotalFootball.ai extracts 17 skeletal keypoints per player from standard footage at 25fps. A football-trained language model interprets those pose sequences and delivers real-time coaching intelligence for match analysis, player recruitment, training, and live commentary.
All players on the pitch are detected simultaneously in every frame, across the full playing area.
Automatic team assignment from kits. Players and officials identified instantly, no configuration required.
17-keypoint skeleton overlay per player. Shoulder angle, hip line, and body orientation extracted on every detected frame.
The Limits of Event Data
Most football analytics platforms are built on tagged event data: discrete actions extracted from video at specific moments. These systems record that a pass happened, not how the passer's body was oriented, nor the spatial context that made it possible or prevented something better.
How It Works
Broadcast or training footage is processed by a convolutional vision model that detects every player on the pitch in each frame, at up to 4K resolution. No specialist cameras or pitch-side hardware required.
17 keypoints are detected per player (nose, shoulders, elbows, left hip, right hip, knees, ankles) and connected to form a real-time skeleton representing exact body posture and orientation.
Structured pose data is passed to the Total Football LLM, a domain-specific model trained on 100,000 hours of annotated match footage, coaching sessions, and tactical analysis. It translates keypoint sequences and orientation patterns into natural-language observations a coaching team can act on.
The Language Model
The Total Football LLM is a domain-specific language model trained on annotated match footage, training sessions, and coaching commentary, totalling approximately 100,000 hours of material.
It is fine-tuned to interpret sequences of skeletal keypoints and translate spatial and orientation patterns into natural-language observations, framed in the vocabulary coaches and analysts already use.
Applications
Body orientation, spatial positioning, and movement data are extracted continuously throughout the match. The LLM layer translates these into coaching observations, delivered in real time.
Skeletal keypoint data extracted from existing footage provides objective measurements of a target player's body mechanics, positional tendencies, and movement signature, independent of team or match context.
Real-time pose data drives immediate audio feedback via Bluetooth earphones, delivered in the player's preferred language. Corrections and movement cues are issued during the session based on live skeletal data, not reviewed from video afterwards.
Analysis
Pose estimation and LLM analysis combine to produce a broad set of outputs across fitness, tactics, individual technique, and team shape.
Distance, speed, and acceleration data captured at 25fps and segmented by intensity threshold. Covers every player across match and training footage.
Sprint count, peak speed, distance, and recovery time per player, segmented by position and game state. Identifies fatigue patterns across a session or match.
Automatically detects and categorises attacking and defensive runs away from the ball — overlapping runs, third-man runs, channel runs, and diagonal movements.
Measures how quickly teams and individuals reorganise following a change of possession, in both attacking and defensive directions, from first reaction to shape recovery.
Identifies each team's true positional shape in and out of possession. Tracks defensive line height, width, structural compactness, and how shape shifts through transitions.
Breaks each possession into build-up, progression, and creation phases, with full positional data for all 22 players throughout each phase.
Identifies the positional and postural cues that initiate a team's coordinated press. Measures the speed, compactness, and effectiveness of each trigger event.
Tracks ground covered by each player during out-of-possession phases, identifying defensive workload distribution and structural gaps in the team's defensive shape.
Shoulder angle and hip line extracted per frame. Captures whether each player is open or closed relative to goal, teammates, and the defensive shape at the moment of receipt.
Body posture at the moment of receipt determines a player's ability to turn and progress. Every receiving situation is tracked, categorised, and scored.
Body orientation, balance, and weight distribution in 1v1 situations. Identifies whether a player is set to win possession or being manipulated out of position.
Detects and quantifies when defenders close down a player in possession — speed, angle, and body position of the press captured for every pressing situation.
Identifies passes that successfully break a defensive or midfield line, using real-time player position data to define where each line sits in every frame.
Player orientation and positioning at the moment of passing, combined with teammate locations, to quantify pass difficulty and quality of option selection.
Player locations, movement patterns, and body orientation on every set piece, from delivery through to second ball. Covers corners, free kicks, and throw-ins.
Identifies all ball-carrying runs, tracking duration, distance, direction relative to goal, and the defensive shape encountered during each carry.
Deployment
The Total Football LLM is available in two deployment configurations. Both run the same underlying model and produce identical output. The difference is where inference happens.
No connectivity required
Inference runs entirely on local hardware at the venue. No data leaves the facility and no network connection is required, making it suitable for training pitches, pre-season camps, or away fixtures where connectivity cannot be guaranteed.
Always current, always scalable
Inference is handled by our hosted infrastructure. The model receives continuous updates as new training data is incorporated, and capacity scales automatically with session volume. Connects directly to existing analytics pipelines via REST or WebSocket.
Integration
TotalFootball.ai outputs structured data in real time. A REST API handles live streaming pipelines; standard formats (JSON and XML) support existing analytics tools; a native MCP server enables integration with AI-native toolchains.
TotalFootball.ai has no proprietary interface. It is a pure intelligence layer: all output is consumed through the API, leaving your team free to build on top of it using any visualisation tool, analytics platform, or custom application.
TotalFootball.ai exposes a native Model Context Protocol (MCP) server. MCP is an open standard that allows AI models and agents to query live data sources directly. This means external LLMs and AI assistants can query pose data, player orientation, and LLM analysis in real time, without custom adapters or middleware.
We're onboarding a limited number of clubs in the initial access phase. Submit your details and we'll be in touch to arrange a technical demonstration.
No commitment. Response within 48 hours.
Commentary
Live positional and pose data feeds a real-time commentary engine that generates spoken narration and tactical analysis from objective spatial data, in multiple languages, suitable for television, radio, or digital broadcast.