Bat-ball contact event detection from audio, not video. The only event-detection method in the production stack that survives noisy training environments where visual methods fail — nets, occlusion, motion blur, partial bat-on-ball contact. ~85% production accuracy, fine-tune peaked at 86.36%, built in a 9-day arc with a memorable annotation-export bug in the middle.
Most amateur cricket and badminton happens in nets. Indoor nets, outdoor nets, mixed-lighting nets, single-camera handheld phone-on-tripod nets. This is where the user shoots their training video — and this is exactly where the visual-only event-detection pipeline starts to fall apart.
In broadcast footage, the ball-on-bat contact frame is unambiguous: clean light, stable camera, 4K resolution, the ball is always visible. In nets footage, the same event is partially occluded by the net itself, partially clipped by the camera framing, partially blurred by 30 FPS motion, partially happening behind the batter's body. The visual pipeline reports either no detection (model didn't see it) or false detection (model saw the ball pass and assumed contact). Either way: analytics downstream gets corrupted.
The audio modality is invariant to all of this. The contact sound is the contact sound — same acoustic signature regardless of camera angle, lighting, net density, or motion blur. If we can train a model to classify the bat-ball contact sound vs ambient noise, we have an event detector that works in the same place the user actually generates the footage.
The alternative would have been investing more in visual robustness — better occlusion handling, motion-blur compensation, multi-frame temporal smoothing. We chose audio. Here's the comparison that drove the decision.
The visual pipeline that powers shuttle tracking and cricket ball tracking also tries to detect bat-ball contact. In broadcast footage it's fine. In nets footage it isn't.
The acoustic signature of bat-ball contact is modality-invariant. The microphone hears the contact sound whether or not the camera saw it.
The build cycle spanned 9 days end-to-end (Jun 19–27, 2025), but the actual accuracy jump happened in a single overnight cycle — after an annotation-export bug was identified and fixed.
I have annotated the clips twice, but when exporting in JSON format, every annotation is marked as miss.
yash.v@khiladipro.com · Linear AI-68 comment
This is the kind of bug that doesn't break anything visibly. The LabelStudio UI showed both hit and miss annotations correctly. The training pipeline accepted the exported JSON without error. But the exported labels were corrupted — every annotation, whether marked hit or miss in the UI, exported as "miss." A model trained on that data would learn to predict "miss" every time, regardless of the actual audio.
The bug was in the LabelStudio JSON export configuration — a templating issue where the hit-label keyword wasn't being written to the exported file. The fix took less than an hour once identified. Finding it took longer.
Why this matters for the portfolio: the lesson isn't "labelling tools have bugs." The lesson is silent data corruption is the most dangerous bug class in ML pipelines. If the bug had survived undetected, every downstream training cycle would have been chasing model improvements against corrupted ground truth — a model that "improves" without ever working. The willingness to annotate twice and notice the discrepancy, dig into the export pipeline, and find the real story is what unlocked the 86.36% the next morning. Most ML projects die quietly to silent data corruption. This one didn't.
By 13:35 UTC the same day, the training pipeline was re-run on corrected labels: Total samples loaded: 106, Model Accuracy: 0.68. By 13:37 UTC: "The accuracy is 68% I will work on it further to improve the accuracy." By 08:42 UTC the next morning (Jun 27): "Improved Model Accuracy: 86.36%".
The full sequence from annotation pipeline setup to production deployment. Includes the bug, the fix, the fine-tune peak, and the September dataset expansion that stabilized the model at ~85%.
The production model serves as the event-detection layer for the KhiladiPro analytics pipeline on training-environment footage. Where visual contact detection drops to unusable, audio detection holds. This is the model that lets KhiladiPro work where the user actually trains — nets, indoor courts, single-camera phone setups — instead of being limited to broadcast-quality footage.
The ~85% production number is the deliberately-honest framing. The fine-tune peak hit 86.36% on the original 106-sample dataset; production on the expanded dataset is slightly lower because the expansion deliberately added harder environments. Reporting the lower number that's more representative of real-world performance is the more credible move than reporting the higher number from a smaller, easier dataset.
Every system has limitations. Pretending the audio detector is bulletproof would be dishonest — and would set up downstream consumers for failures they couldn't anticipate.