Sports Video Intelligence · Hangzhou, China

From match footage to
tactical insight.

LynxAct combines vision-language models, multi-object tracking, and LLM reasoning to automate sports video analysis — from annotation pipelines to coach-ready tactical reports.

View on GitHub See it in action
98.7%
zero-shot VLM accuracy
on activity recognition probe
±1 m
pitch calibration
reprojection accuracy
4-tool
function-calling loop
in the AI Coach Agent

Products

One pipeline, three layers

We build open tooling for every layer of sports video understanding — so clubs, analysts, and researchers can adopt exactly the depth they need.

Tactical Annotation

VLM-assisted pre-annotation with human expert review. Annotators verify machine proposals instead of labeling from scratch — cutting manual tagging time by an order of magnitude while keeping expert judgment in the loop.

VLM pre-labeling · Expert verification · Event taxonomy

AI Coach Agent

An LLM agent that turns match data into natural-language tactical reports. A four-tool function-calling loop queries events, surfaces patterns, and drafts analysis a coach can act on — backed by a curated library of 62 technique cards.

LLM reasoning · Function calling · Tactical reports

Player Tracking

An open pipeline that extracts 2D pitch coordinates from broadcast footage — detection, multi-object tracking, and pitch calibration. Validated on professional match footage with sub-meter reprojection accuracy.

YOLO · ByteTrack · Homography calibration

Technology

Measured, not promised

Every capability on this page is backed by a reproducible probe on real data. Here is what we have verified so far.

98.7%Zero-shot VLM accuracy on a 150-image activity-recognition probe (Stanford40), 8 of 10 classes perfect
0.62–1.25mPitch-calibration reprojection error on professional broadcast footage, 5-anchor homography
95.7%Of frames with ≥8 tracked players — 12.84 players per frame on average, YOLOv8 + ByteTrack
62Technique cards in the Coach Agent knowledge base, covering dribbling, passing, and finishing
Match footage
Detection & tracking
Pitch calibration
Tactical events
LLM analysis
Coach report

In action

Working software, not slideware

Pre-annotation studio. Three prediction sources are fused into a single proposal; the expert confirms or corrects with one click.
AI Coach Agent output: automatically identified top moments from a match
Coach Agent output. The agent scans match data and surfaces the moments that matter, with reasoning attached.

Open source

Built in the open

Our core pipelines are Apache-2.0 licensed. Star them, fork them, or ship them into your own stack.

lynxact-coach Apache-2.0

AI Coach Agent for football tactical analysis — a four-tool function-calling loop over match data, producing natural-language coaching reports.
lynxmove-oss Apache-2.0

Multi-source fusion pre-annotation toolkit for sports video — VLM proposals, player-prior signals, and expert review in one web UI.

Company

Built by practitioners

LynxFlow AI

LynxAct is developed by LynxFlow AI, a team founded in 2026 and based in Hangzhou, China. We build AI tooling for sports analysis — annotation pipelines, tracking infrastructure, and LLM-powered reporting.

Founded
2026
Headquarters
Hangzhou, China
Focus
Sports AI · Video intelligence
Stage
Bootstrapped, pre-seed

Founder

Alexander Cardoza

Founder & CEO, LynxFlow AI

Alexander holds dual bachelor's degrees (B.Ed. and B.A.) from a 211 university in central China, and an M.Sc. from the University of Edinburgh, with research focused on accelerometer data from smart wearable devices and physical-activity epidemiology. He started LynxAct to bring modern vision-language and LLM tooling to a field still run on manual tagging.

Reach us at [email protected]