Break Point

Inspiration Table tennis is the fastest sport on earth. It is mostly mental, yet almost all players train with no data and no real feedback. Professional players have analysts tracking every spin, bounce, and hesitation. Everyone else just records a video and guesses what went wrong. After a match, advice usually sounds like this: Be more aggressive Move your feet Stop pushing so much

But when did that happen Why did it happen Against who does it always happen

Now imagine this. You played someone a year ago and still have that video. You also have their recent match and your recent match. No human coach can watch all of that, remember everything, and connect the patterns. Break Point can. We did not want another score or highlight app. We wanted a system that understands why players lose and tells them how to win next time. What It Does Break Point turns raw match footage into real coaching in minutes. It does not just track shots. It understands patterns. Break Point detects what shots you play, how long rallies last, where the ball is placed, how your footwork changes under pressure, and when you hesitate after defensive shots. It compares your past matches, your opponent’s past matches, and how you play against that specific opponent. From this, it produces the real problem, a match specific strategy, and one mental cue you can remember during the game. It is not a dashboard. It is a coach that speaks.

How We Built It

Break Point is an end to end system that separates seeing the match from understanding the match.

First, we built a computer vision pipeline that works with real phone footage. The system detects the table, tracks the ball at high speed, and follows player movement even when the camera moves. From this, we extract raw events like hits, bounces, placement, and recovery movement. Next, we turn those events into meaningful metrics. We calculate footwork speed, footwork drop under pressure, arm speed, rally length, shot distribution, and a custom aggression index. Aggression measures intent, not just speed, using shot depth and court position. Each match is stored as structured data in Snowflake. To give the system memory, we generate a semantic summary for every match and convert it into a vector using Snowflake Cortex. This allows the system to remember how a player behaves across matches. When a new match is uploaded, we retrieve similar past matches and send both the current match and historical context into Kimi K2 Think. The model reasons step by step, connecting physical changes like footwork slowdown to mental outcomes like hesitation and passive play. Finally, we translate everything into coaching output a player can actually use. Break Point gives one clear problem, one concrete fix, and one short mental cue. Using ElevenLabs, the coach speaks directly to the player.

Challenges We Ran Into Tracking table tennis from phone footage is hard. The ball is small and fast, lighting changes, and camera angles vary. Early versions produced too many stats without clear coaching value. We had to remove anything that did not help the player win the next match. We also had to prevent AI hallucinations by forcing the reasoning model to only use the extracted match data.

Accomplishments We Are Proud Of We built a system that remembers and compares matches across time. We turned complex movement and behavior data into advice players can use mid match. We combined computer vision, reasoning models, vector memory, and voice into one product. What We Learned More data does not mean better coaching. Mental pressure shows up clearly in physical movement. One clear cue is more valuable than many statistics.

What Is Next for Break Point We want to support full match and season long analysis. We plan to add real time feedback using on device processing. We want to integrate wearables to understand physical and mental pressure together. Long term, Break Point becomes a personal match memory for athletes so they walk into every match prepared to win.

Built With Kimi K2 Think Gemini Flash Snowflake and Snowflake Cortex OpenCV and MediaPipe ElevenLabs React, Vite, FastAPI, Python