At Samsung Research America, I built a “talk to your TV” assistant that uses recent video frames to answer questions about what’s playing. I also added persistent watch queries that monitor a live feed and send alerts until stopped.
On DGX Spark, the alerts fired in 0.6 seconds versus roughly 10 seconds for the NVIDIA VSS 3.3 benchmark, using about 58% fewer input tokens and catching all 21 test events with zero false alarms. I also built a confidence-gated cascade to route queries between small and large models.
At NVIDIA, I tuned PyTorch and TensorRT inference performance with ML, compiler, and deployment engineers, cutting end-to-end latency by roughly 18% on targeted workloads. I profiled models with Nsight Systems and analyzed CUDA kernels to identify bottlenecks and evaluate quantization trade-offs.
At Airtrek Robotics, I built a synthetic data generation pipeline that reduced dataset creation cost by 60% and improved YOLO accuracy by 20% on edge cases. I also deployed perception and motion-aware inference pipelines on NVIDIA Jetson Orin NX, reaching about 20 ms latency.
