At the intersection of human kinetics and artificial intelligence lies a new frontier of understanding. Willow Cortex is not just an API; it is a steerable interface to a Realtime World Model. It is a system thoughtfully designed to reconstruct physical reality from video, built on a foundation of safety, transparency, and unparalleled performance.
How It Works: A Steerable, Developer-First Process
Our process is designed for precision and control. It begins when a partner's application makes a secure request to a serverless function, receiving a temporary signed URL in return. The video file is then transferred directly to a cloud storage bucket, accompanied by a metadata packet. This metadata acts as the "prompt" for our World Model, allowing you to trigger our full processing pipeline, select individual analysis modules a la carte, or define a custom combination of services. This ensures you have steerable, granular control over the analysis from the very first step. The final output of every service is delivered as a single, validated JSON payload for seamless integration.
The Engine: Unparalleled Performance at Scale
The engine for this analysis is a 100% serverless environment. When your video is received, our asynchronous functions spin up instantly and operate in parallel, creating what is by far the most performant solution in the world for this level of analysis. This architecture ensures massive scalability and efficiency, processing complex biomechanical tasks with incredible speed. The resulting data - a rich tapestry of biomechanics, kinetics, and contextual metrics - is organized within a sophisticated backend of schemaless NoSQL databases, vector databases for semantic understanding of motion, and knowledge graphs to map complex relationships.
From Analysis to Exploration: Unlocking New Potential
Decoding motion is only the beginning. Our platform allows you to move beyond analysis and into exploration. By applying an ensemble of probabilistic models we can simulate the athletic motions we've decoded. For instance, we can analyze every swing from a batting cage session and simulate thousands of potential outcomes to generate predictive statistics. This same principled approach can be used to model the probability of future injury across different time periods, providing a more holistic understanding of an athlete's physical load and risk. It's about using AI not just to see what happened, but to responsibly explore what could happen. This capability unlocks a vast spectrum of applications, empowering everything from next-generation remote coaching and data-driven broadcast media to immersive fan gamification, objective talent scouting, injury prevention analytics, and metaverse-ready motion capture.
Contact us in the chat or submit a ticket to learn more.
Master Service Catalog for Metadata-Triggered Analysis
I. Foundational Triage, Perception & World Orientation
These services establish the AI's "visual cortex," orienting the World Model in physical space and time before measurement begins.
video_quality_assessment: Determines if the video is of usable quality (lighting, obstructions, clarity) for accurate biomechanical reconstruction.
sport_detection: Identifies the primary sport being played (e.g., "golf", "baseball", "softball", "tennis", etc.) to load the correct physics models.
physics_based_handedness_detection: Deterministically identifies handedness (Left/Right) by comparing the 3D peak velocity of limbs. This bypasses visual occlusion issues by tracking energy transfer and ballistic intent.
view_axis_determination: Uses "Aspect Ratio of Motion" logic to mathematically determine if the camera is viewing the action from a "Side" (Traversing) or "End" (Tunnel) perspective, independent of distance.
camera_angle_classification: Refines the axis determination into specific coaching views (e.g., "Down-the-Line", "Face-On", "Open Side View", "Catcher View") using geometry at the "Set Position."
video_structure_analysis: Classifies the video's format (e.g., "single_rep", "multi_rep_session", "gameplay_footage", "highlight_reel").
action_segmentation: A temporal segmentation service that locates and time-stamps every primary athletic action (swing, pitch, throw), creating a list of analyzable instances.
temporal_landmark_detection: Identifies the precise frame indices for critical kinetic states: The "Set" (Minimum Kinetic Energy), The "Kinetic Peak" (Max Velocity), and "Impact/Release."
phase_visibility_assessment: Based on the camera angle, determines which specific biomechanical phases (e.g., "Takeaway", "Arm Action") are geometrically visible and which are occluded.
environment_detection: Analyzes the scene to identify the setting (e.g., "indoor_tunnel", "outdoor_range", "on_field", "backyard").
equipment_detection_and_branding: Identifies key equipment (bats, clubs, gloves, shoes) and, where possible, their manufacturer branding.
apparel_detection_and_branding: Identifies apparel (hats, shirts, pants) and their associated brands.
II. Core Vision & Raw Data Extraction
These services extract the fundamental, frame-by-frame training data that powers the World Model.
dense_pose_estimation_2d: Extracts the (x, y) pixel coordinates for all major body joints for every frame of a segmented action.
dense_pose_estimation_3d_inferred: Infers a z-coordinate (depth) for each joint to create a 3D representation of the athlete's pose in space.
end_effector_tracking: Provides a dedicated, high-fidelity time-series of coordinates for the primary implement's point of interest (club head, bat barrel, throwing hand).
ball_flight_tracking_2d: Tracks the ball's trajectory from the moment of impact or release until it leaves the frame.
calibration_method_detection: Automatically detects the best available calibration source (ArUco Sheet, Known Objects, or Anthropometric Scaling) and returns the confidence level.
scene_scale_calibration: Returns the calculated Homography Matrix or Scale Factor used to convert pixels to real-world units (inches/meters).
video_fps_estimation: Provides an estimated frames-per-second count for the source video, critical for temporal calculations.
audio_event_detection: Pinpoints timestamps of key audio events, such as bat-on-ball or club-on-ball impact, to synchronize video and audio data.
speech_to_text_transcription: Transcribes any spoken words from the audio track into searchable text.
optical_character_recognition_ocr: Detects and transcribes any text or numbers visible in the video, such as data from on-screen graphics or launch monitors.
event_counting: Counts the total number of occurrences of a specified event (e.g., "number of pitches thrown").
III. Biomechanical & Physics Engine
This is the core scientific analysis layer, using our proprietary physics engine to turn raw coordinates into actionable, calculated truths.
A. Path & Trajectory Analysis (Topology)
path_smoothness_analysis_jerk: Calculates the "jerk" (rate of change of acceleration) of the end-effector to quantify motion smoothness and efficiency.
trajectory_efficiency_analysis_curvature: Calculates the curvature of the end-effector's path to identify inefficient loops or mechanical deviations.
path_planarity_analysis_torsion: Uses geometric torsion calculations to determine how strictly a swing or throw adheres to a single plane. (from DaVinci engine)
B. Body Segment Dynamics (Kinematics)
center_of_mass_estimation_3d: Calculates the 3D trajectory of the athlete's Center of Mass (CoM) to analyze weight transfer, stability, and ground interaction.
kinematic_sequencing_analysis: Measures the peak angular velocity of each body segment (pelvis, torso, arm) and their relative timing to verify an efficient proximal-to-distal energy transfer.
segmental_angular_velocity_measurement: Provides the peak rotational speed (in deg/s) for individual body segments.
rotational_separation_analysis_x_factor: Calculates the maximum angle of separation between the hips and shoulders (X-Factor), a key source of potential power.
C. Forces & Power (Kinetics)
estimated_power_output_kw: Synthesizes kinematic data to estimate the peak power output (in kilowatts) of the athlete's motion.
ground_reaction_force_estimation_grf: Analyzes vertical acceleration of the Center of Mass to estimate the peak force the athlete generates from the ground interaction.
D. Computational Engine Analysis
motion_rhythm_and_tempo_analysis: Uses Fast Fourier Transform (FFT) on motion data to analyze the dominant tempo and rhythmic consistency of a movement. (from DaVinci engine)
motion_compactness_analysis: Calculates the 3D convex hull volume of the entire motion path to quantify its spatial efficiency and compactness. (from DaVinci engine)
inter_rep_stability_analysis: Compares the motion paths of multiple repetitions to generate a "repeatability score," quantifying the athlete's mechanical consistency. (from DaVinci engine)
E. Corrective Analysis
inverse_kinematics_corrective_delta: "Works backward" from a theoretically optimal impact/release position to calculate the specific, quantitative changes in joint angles (e.g., "+10 deg hip rotation") required to achieve it.
IV. Performance Metrics & Outcome Analysis
These services derive high-level, sport-specific Key Performance Indicators from the biomechanical data.
action_outcome_analysis: Determines the result of the action (e.g., "ball in play", "fair ball", "swing and miss", "shot shape: draw").
hitting_metrics_estimation: Provides calculated values for metrics like Exit Velocity, Launch Angle, Bat Speed, Attack Angle, and more.
pitching_metrics_estimation: Provides calculated values for metrics like Pitch Velocity, Spin Rate, Release Height, Extension, Arm Slot, etc.
golf_swing_metrics_estimation: Provides calculated values for metrics like Club Head Speed, Ball Speed, Smash Factor, Spin Rate, Carry Distance, etc.
sport_specific_kpi_suite: A master trigger to run the full, relevant list of metrics for the detected sport (combines the three services above).
V. Predictive & Simulation Engine
These advanced services allow the World Model to project future states and risks based on current data.
statistical_outcome_simulation: Takes data from a single session (e.g., batting practice) and uses an ensemble of probabilistic models to simulate a full season of statistics (e.g., projected batting average).
injury_risk_modeling: Analyzes kinetic and kinematic data to calculate joint load and stress, modeling the cumulative risk of injury over a session, week, or season.
fatigue_detection_analysis: Monitors biomechanical markers over the duration of a session to detect degradation in mechanics, indicating the onset of fatigue.
performance_forecasting: Uses time-series analysis on an athlete's historical data to forecast potential improvements or regressions in key performance metrics.
VI. Content & Report Generation
These services package the World Model's insights into user-facing products and media.
generative_coaching_report: Synthesizes all data into a complete, AI-written coaching report with ratings, explanations, and actionable advice.
pro_player_comparison: Compares the user's key biomechanical signatures to those of a professional athlete in our database.
corrective_drill_recommendation: Suggests specific training drills based on the identified mechanical flaws.
automated_video_clipping_service: Generates a JSON map of start/end times for every biomechanical phase (Takeaway, Downswing, etc.) for use in automated video editing.
automated_highlight_reel_generation: Identifies the reps with the highest performance metrics (e.g., top exit velocity) and generates clips for social sharing.
avatar_motion_mapping_3d: Exports the dense 3D pose data in a format (e.g., FBX, BVH) suitable for animating a 3D character or avatar in a game engine or metaverse environment.
virtual_shot_overlay: Creates a data overlay for the video, showing the ball's flight path, key metrics, and other graphics in augmented reality.
grounded_analysis_chat_session: (New) Enables an interactive chat experience where a user can ask natural language questions about their results. The AI is grounded exclusively in the data from the specific analysisId, allowing it to provide contextually-aware answers (e.g., "Explain my kinematic sequence in more detail" or "Why was my power score lower on the third rep?").
VII. Longitudinal & Profile Management
(These services enable the tracking of performance and biomechanics over time for a specific individual or cohort.)
progress_tracking_suite: A comprehensive service for managing and analyzing an athlete's journey.
Action: create_progress_tracking_profile: Initializes a new longitudinal profile for an athlete or group. This service takes a profileName (e.g., "John Doe - 2025 Season") and returns a unique profileId.
Action: add_analysis_to_progress_profile: Appends a completed analysisId to a specified profileId, linking that session's data to the athlete or cohort's historical record.
Action: retrieve_progress_profile_data: Retrieves the complete, time-series dataset for a given profileId. This returns all associated analyses, allowing partners to plot key biomechanical signatures and performance metrics over time to visualize trends, improvements, or regressions.