At WAYMO, I migrated legacy perception models from TensorFlow 1.x to PyTorch 2.0, validating output parity and reducing inference latency by approximately 30%. I also developed object detection and semantic segmentation models for dense urban scenes.
I built Apache Beam and Google Cloud Dataflow pipelines for LiDAR and camera data, reducing dataset preparation time by 40%. For custom TPU deployment, I optimized detection models through quantization and graph optimization, increasing throughput by 25% with less than 0.5% mAP loss.
At LTIMINDTREE (Client: Honeywell), I developed predictive-maintenance and anomaly-detection models, and built a RAG prototype for maintenance documentation. At GE HEALTHCARE, I developed an MRI diagnostic tool and 3D CNNs for CT segmentation, and deployed inference APIs with monitoring.
