InCoach
Golf live trainer which collects, researches, compiles and manages drills and exercises to improve your game.
Core trick
Runs MediaPipe pose tracking live and compares the golfer's positions against drill reference frames from parallel video-analysis agents.
Description
Golf live trainer which collects, researches, compiles and manages drills and exercises to improve your game. Uses dispatched research agents to find drills from creators you like and converts them to a blueprint for a guided session.
Validates claims/drills in the video using parallel agents to avoid shoddy claims in videos. Specifies important moments in the videos as dynamic visual aid for the session.
MediaPipe tracks your positioning and interfaces with the live coaching agent, evaluating your performance vs the grill and guides you through it. Responds to your feedback, saves your session so you can improve over time without needing to manually track your position, status and goals.
README excerpt
# InCoach - AI Visual golf coach
<img width="1672" height="941" alt="image" src="https://github.com/user-attachments/assets/a7240252-8b77-43d3-a9d5-1bcc055e715d" />
Start with what you want to improve on ("fix my slice in 20 min, sore back") and a multi-agent pipeline compiles a researched, safety-validated drill routine with key-moment video chapters.
## Demo flow
1. **Start with intent.** Hold to talk, or type. The UI streams each agent's reasoning live: Researcher's search queries, Validator's safety diffs, parallel VideoScout / MomentMiner workers labeled `#1`, `#2`, `#3` as they finish.
2. **Routine appears.** Each drill is an embedded player with captioned moment chips. Click a chip → seek to that exact frame.
3. **Open the Live Coach.** Hold to talk. *"Show me the follow-through."* The UI jumps; the coach narrates.
4. **Practice on camera.** A skeleton overlays your video. `swayScore` tracks posture against the drill in real time.
## Description
**The compile pipeline.** A request flows through a chain of Gemini Managed Agent interactions: `IntentParser` extracts goal/skill/constraints → `Researcher` grounds in Google Search → `Compositor` drafts a structured routine → `Validator` runs a safety pass that returns explicit `before / after / reason` diffs → a fanout of `VideoScout` sub-agents picks one YouTube demo per drill → `MomentMiner` sub-agents watch each video in parallel and emit 2–4 timestamped key moments. Every stage streams `AgentEvent`s to the client