Salman Asad

- Designed and deployed an agentic AI system using LangChain & LangGraph, enabling automated mapping of music metadata through worker–supervisor LLM loops. - Built a Natural Language Music Search feature to search for tracks using description of music. - Enhanced Triton Inference Server configuration, reducing GPU usage by 30–40%, allowing multiple models to run on the same hardware. - Developed scalable FastAPI microservices to integrate LLMs, embeddings, and retrieval components into production systems. - Evaluated ElasticSearch and Qdrant for high-scale vector retrieval, contributing to architectural decisions for future AI infrastructure. - Improved prompt design, model selection, and evaluation criteria to boost LLM output consistency and relevance. - Collaborated with leadership to iterate on AI roadmap items, focusing on agent design, LLM optimization, and workflow orchestration.

- Learned and applied techniques to convert machine learning models to ONNX and TensorRT, improving inference speed and efficiency for real-time applications. Explored and implemented SAHI Inferencing and Anomalib for visual anomaly detection to enhance understanding of quality control solutions. - Worked on implementing object detection and segmentation algorithms, such as RT-DETR, YOLOv10, SOLOv2 and more, to strengthen the visual quality inspection system. - Gained practical experience in training and testing machine learning models for real-time visual quality inspection, ensuring the system meets accuracy and reliability requirements.

I worked on a variety of tasks within the Strada Imaging's DeepRoad ecosystem, including: - Processed and annotated datasets for segmentation models used in autonomous road assessment. - Image processing-based detection for a hard-to-segment surface type in road images - Applied classical CV and anomaly detection methods to enhance detection accuracy.