Open to opportunities

Salman Asad

@salmanasad

Message

Data Scientist specializing in LLM-powered audio search and production MLOps across NLP and computer vision.

Pakistan

Message

What I'm looking for

I’m looking for a role where I can build production AI systems—agentic RAG, multimodal ML, and efficient inference—owning end-to-end MLOps and performance benchmarking with a team that values reliability and iteration.

I’m a Data Scientist who builds multimodal AI systems that move from research to production. I’ve trained and deployed audio models for vocal/instrumental classification and music taxonomy, with experiments tracked end-to-end in MLflow.

In my current role, I trained a CRNN for vocal/instrumental classification (90% accuracy, F1 of 89) using MIL-based spectrogram aggregation and augmentation on 60k tracks. I also fine-tuned an Audio Spectrogram Transformer to a custom 1,343-tag music taxonomy (Jaccard Score 79%, Hamming Loss 0.002, NDCG Score 92%), enabling stronger retrieval and ranking.

I focus heavily on real-world performance and inference efficiency. I evaluated and deployed DeepSeek-R1 across multiple backends (VLLM, Ollama, Llama.cpp, SGLang) using provider abstraction patterns, and optimized Triton Inference Server to reduce GPU usage by 30–40% for multi-model deployment.

I also design agentic systems that produce reliable, structured outputs. I built an LLM-powered agentic RAG for natural-language music search with FAISS and FastAPI (async/await + Pydantic validation), and I’ve developed multi-agent LangGraph workflows for mapping music metadata across taxonomies using confidence scoring and robust fallback handling.

Experience

Work history, roles, and key accomplishments

Current

Data Scientist

Current

Upright Music/CEBS

Jul 2024 - Present (1 year 11 months)

Trained a CRNN for vocal/instrumental classification on 60k tracks, achieving 90% accuracy and F1 of 89%, with full experiment tracking in MLflow. Fine-tuned an audio spectrogram transformer for a 1,343-tag music taxonomy across 140k tracks (Jaccard 79%, Hamming loss 0.002, NDCG 92%) and built an LLM-powered agentic RAG music search service deployed via FastAPI.

Librosa MLFlow DeepSeek R1 fastAPI LangGraph Deepseek R1

Computer Vision Engineer

Axis.ai

Jun 2021 - Jul 2024 (3 years 1 month)

Converted and optimized models using ONNX and TensorRT for real-time inference, improving deployment reliability via latency/throughput evaluation and hyperparameter tuning. Researched anomaly detection frameworks (Anomalib, SAHI) and built high-throughput detection systems with YOLOv10 and RT-DETR for quality inspection and defect pattern identification.

Onnx TensorRT YOLOv10 RT DETR Anomaly Detection Computer Vision Optimization

Computer Vision Engineer

Strada Imaging

Mar 2020 - Jul 2021 (1 year 4 months)

Prepared and annotated large-scale segmentation datasets to support high-quality training inputs. Applied classical computer vision methods for anomaly detection and feature extraction to iterate on model performance.

Segmentation Data Annotation Anomaly Detection Feature Extraction