At Blaize, Inc., I built a YOLOv8 aerial object detection system for video analytics on Blaize edge hardware. Error analysis and a small-object detection head raised mAP50:95 from 64.4 to 66.2.
I also distilled a YOLOv5x detector into a YOLOv5m student, cutting parameters by about 75% and compute by about 4x while keeping accuracy within 1 to 2 mAP points. For 2K video, I built an adaptive tiled inference pipeline using SAHI to balance detection accuracy and latency.
For video-frame analysis, I evaluated LLM serving options and deployed Qwen3-VL-30B-A3B across multiple GPUs with an image-query API. I also rebuilt the detection and segmentation auto-labeling pipeline on Grounding DINO and SAM, reducing manual labeling effort by 20%.
At Mythic, Inc., I adapted object detection models to analog in-memory compute hardware using noise-aware training, quantization-aware training, and pruning. Earlier at Ignitarium Technology Pvt. Ltd., I deployed digit recognition on a Renesas MCU and optimized a footwear defect detection pipeline with TensorRT.

