Drop a football clip into a folder. Master Coach watches it like an elite analyst, researches the match, and turns it into an animated tactical board with coach fixes.
Master Coach is a local-first AI tactical studio for football coaches. Point it at a video clip and it produces a synced video player, an animated 2D strategy board, broadcast-frame overlays drawn by Gemini Nano Banana, key-moment timelines, and corrective "coach fix" counter-plays — all from a single click.
Built for the GDG Newport Beach × RocketRide × GMI Cloud × NVIDIA hackathon.
You drop an .mp4 into clips/inbox/. The app picks it up, and on Launch Analysis it kicks off a 10-stage pipeline that:
- Reads the clip and extracts frames with ffmpeg
- Researches the match context on the web (Google Search grounding)
- Builds a structured match-facts JSON (teams, lineups, numbers, goals)
- Sends the clip into Gemini for full multimodal tactical understanding
- Classifies every frame for key-moment density (Google or GMI Cloud)
- Picks the best hero frames and asks Nano Banana to draw tactical arrows on them
- Generates a normalized 2D strategy scene
- Generates a coach-fix counter-play layer
The result is rendered in a synced three-pane analysis view.
A custom SVG pitch renders players as numbered chips, passes and runs as arrows, and space-creation as gold zones. It is keyframed against the video timeline so scrubbing the clip moves the board, and selecting an event jumps both.
Toggle Coach Fix mode and a second violet overlay drops in — corrective runs, alternate passing lanes, and the training drill that drills the pattern.
For each hero moment, Gemini's gemini-2.5-flash-image-preview model paints cyan arrows, violet runs, and gold space-zones directly onto the broadcast frame. No green-screen, no manual annotation — just the real broadcast image with tactical overlays.
Every stage of the RocketRide workflow is visible live as it runs, so the demo never feels like a black box.
| Layer | Tech |
|---|---|
| App | Next.js 15 (App Router), TypeScript, Tailwind |
| Tactical reasoning | Google Gemini gemini-2.5-flash (multimodal video) |
| Broadcast annotation | Gemini gemini-2.5-flash-image-preview (Nano Banana) |
| Frame classifier | GMI Cloud Qwen2.5-VL-7B-Instruct (OpenAI-compatible) — runtime toggle |
| Orchestration | RocketRide pipeline (pipelines/master-coach-analysis.pipe) |
| Video | ffmpeg / ffprobe |
The VISION toggle in the top-right of the landing page swaps the per-frame classifier between Google and GMI Cloud at runtime. Tactical reasoning, match facts, counterplay, and Nano Banana stay on Google.
npm install --legacy-peer-deps
npm run devThen drop a clip into clips/inbox/:
clips/inbox/
argentina-france-2022-108-messi-shot.mp4
argentina-france-2022-108-messi-shot.meta.json # optional sidecar
Open http://localhost:3000, pick the clip, hit Launch Analysis.
Create .env.local:
GOOGLE_API_KEY=...
GEMINI_TEXT_MODEL=gemini-2.5-flash
GEMINI_VIDEO_MODEL=gemini-2.5-flash
GEMINI_IMAGE_EDIT_MODEL=gemini-2.5-flash-image-preview
# Optional — leave GMI_API_KEY empty to lock the toggle to Google
GMI_API_KEY=
GMI_BASE_URL=https://api.gmi-serving.com/v1
GMI_VISION_MODEL=Qwen/Qwen2.5-VL-7B-Instruct
VISION_PROVIDER=googleYou'll also need ffmpeg and ffprobe on your PATH.
If a clip is ambiguous, drop a .meta.json next to it to ground the pipeline:
{
"competition": "FIFA World Cup",
"season": "2010",
"match": "Spain vs Netherlands",
"minute": "116",
"team_focus": "Spain",
"event": "Iniesta winning goal",
"desired_mode": "improve_if_for_us"
}| # | Stage | Notes |
|---|---|---|
| 1 | intake |
ffprobe for duration + basic media info |
| 2 | extractFrames |
ffmpeg @ 4fps → analysis/{clip}/frames/*.webp |
| 3 | matchContext |
Gemini + Google Search grounding |
| 4 | matchFacts |
Gemini structured-output → teams, lineups, goals |
| 5 | geminiTactical |
Gemini video understanding → events with timestamps + pitch coords |
| 6 | classifyFrames |
VisionProvider per-frame classification (GMI or Google) |
| 7 | selectHeroFrames |
Cross-references Gemini events with classifier hits |
| 8 | annotateFrames |
Nano Banana draws cyan/violet arrows + gold zones |
| 9 | strategyScene |
Deterministic transform → renderable scene JSON |
| 10 | counterplay |
Gemini "coach fix" mode |
All outputs land in analysis/{clip_id}/. The landing page shows a Cached badge for clips that already have results — re-running overwrites.
Master Coach/
app/ # Next.js app router
components/ # Strategy engine, video player, pipeline rail, ...
lib/ # Pipeline runner, provider adapters, parsers
pipelines/ # RocketRide pipeline definition
clips/inbox/ # Drop clips here
analysis/{clip_id}/ # Per-clip outputs (JSON + frames + hero images)
public/landing/ # Stadium artwork
docs/screenshots/ # README screenshots
Hackathon prototype. Not for production use.