At Daimler Truck Innovation Center India, I benchmarked detector and stereo-depth model combinations on KITTI frames and recommended a YOLO + BridgeDepth pipeline. It reached 92.4% accuracy with a 3% strict error rate.
I traced depth-estimation inaccuracies to a disparity-range mismatch between benchmark and in-house client data, then cross-validated findings with ROS2 bag and point-cloud data. I also implemented FP16 inference optimization, halving GPU compute load without materially degrading accuracy.
At Intel, I distilled a Real-ESRGAN teacher model into a lightweight PyTorch CNN using combined L1 and VGG perceptual loss. The model achieved SSIM > 0.92 and PSNR of 30 dB on DIV2K while reducing parameter count and computational overhead.
In my projects, I built LLM and RAG applications for code review, study support, and content summarization. They include repository filtering and chunking, document embeddings, grounded question answering, and exportable summaries.
