← Back to the portfolio

Case study · Founding Computer Vision Lead (Intern) · Jan 2025 – Jan 2026 · San Diego

Playar

Playar turns a cricket net into a match simulator. Regular cameras watch each shot, and the software shows how it would play out in a match, with no wearables or sensors on the bat or ball.

I was the head computer vision engineer from the earliest stage, building the proprietary vision software alongside Qualcomm engineers.

playarsports.com ↗ (opens in a new tab)

shortgoodfullpredicted pathbounceball 0.89detecttrackbouncecontactpredictILLUSTRATION
Illustration of the tracking pipeline, not footage or output from the system.

The system

A cricket ball is small and fast, and its most important moment is often hidden by the bat and the batter. To simulate a shot, the pipeline detects the ball, keeps its identity across frames, finds the moment of contact and predicts the trajectory that follows.

  1. DetectBall detection on every frame
  2. TrackMulti-object tracking across frames
  3. ContactPose estimation tied to ball-contact events
  4. PredictPhysics-based post-contact trajectory at 60 Hz
  5. DeployOptimized edge inference

Tech stack: YOLOv11 · DeepSORT · MediaPipe · TensorRT · ONNX Runtime · Python · Kalman filtering

Results

mAP@0.5 on simulated match footage
98%
approximate per-frame latency, a 3× speedup to about 200 FPS
15 → 5 ms
reliability gain on edge-case contact frames
~18%

Monocular 3D ball tracker

A phone-camera edge application that locates ball pitch and bounce from a single view, using monocular 3D lift and physics priors.

It produced shot-by-shot drop-zone heatmaps and per-session reports for coaching staff, and was piloted at batting academies.

Beyond the model

I combined engineering with product and management work, shaping priorities with leadership as the startup grew, and turning feedback from coaches and batting academies into product features.