Sareh Soltani Nejad

Sareh Soltani Nejad

Data Scientist / Machine Learning Researcher

Vancouver, BC, Canada

About Me

I am Sareh Soltani, a Machine Learning Researcher with over 5 years of experience architecting scalable, end-to-end AI systems. Currently, as a GenAI Data Scientist at Mercor Inc., I specialize in Large Language Models (LLMs), building agentic RAG systems designed for multi-step reasoning and tool integration across large-scale unstructured data. Previously, as a Data Scientist at BrainsCan, I leveraged multimodal data to investigate the effect of music perception on brain functionality, while developing AI-driven solutions for pharmaceutical and healthcare applications. Prior to that, I joined the Vector Institute as a Machine Learning Engineer to collaborate on anomaly detection projects. In this role, I implemented machine learning pipelines and predictive models to solve anomaly detection usecases across the banking, insurance, and surveillance sectors.

I earned my MSc in Computer Science from Western University under the supervision of Dr. Anwar Haque, where I developed a weakly supervised video anomaly detection model based on two stream of I3D network. Before joining Western, I completed my B.Sc. in Computer Engineering, focusing on object detection and tracking systems.

My expertise spans Generative AI, Computer Vision, and Large Language Models, with specialized applications across the healthcare, multimedia, and banking sectors. I have extensive experience architecting production-ready AI systems, implementing advanced ML pipelines, and deploying models that drive tangible business value.

Personal email: sarehsoltani.inbox@gmail.com
Work email: ssolta7@uwo.ca

Projects

Medibot: Generative AI Medical Chatbot
Architected and deployed a scalable LLM Medical Chatbot featuring an end-to-end RAG (Retrieval-Augmented Generation) pipeline. The system integrates Hugging Face Sentence Transformers for dense vector embedding generation and Pinecone for low-latency semantic search. It orchestrates OpenAI models via LangChain, implementing custom history-aware retrievers to manage conversational state and reduce hallucinations by grounding responses in verified medical datasets. I engineered the serving layer using Flask and built a production-grade CI/CD workflow using GitHub Actions, automating testing and deployment directly to AWS EC2 infrastructure. Code
Medibot: Generative AI Medical Chatbot
Real-Time ECG Heartbeat Classification
Built a real-time ECG heartbeat classification system using CNNs trained on the MIT-BIH dataset. The model automatically extracts key temporal ECG features and was optimized with Optuna, with experiments tracked via MLFlow for robust performance. Deployed as a scalable REST API on AWS EC2, the system accepts CSV inputs and delivers fast, accurate predictions, achieving 95% accuracy in classifying arrhythmias. This end-to-end solution enables cardiologists and researchers to analyze their own data in real-world healthcare settings, improving clinical diagnosis through efficient, scalable arrhythmia detection. Code
Real-Time ECG Heartbeat Classification
Immune Cell type Annotation from scRNA-seq Blood Data Using scGPT | In collaboration with UHN
Built an end-to-end transformer-based pipeline to classify ~20K blood cells from scRNA-seq data, identifying ~10 immune subtypes using scGPT embeddings, clustering, and marker gene validation. Preprocessing included HVG selection and canonical immune marker validation. Results highlight the potential of large-scale generative models for high-resolution immune profiling. Code
Immune Cell type Annotation from scRNA-seq Blood Data Using scGPT | In collaboration with UHN
Weakly-supervised Anomaly Detection in Surveillance Videos (MSc Thesis)
This project introduces a cutting-edge approach to anomaly detection in urban surveillance systems using Two-Stream Inflated 3D (I3D) Convolutional Networks. By capturing both spatial and temporal features more effectively than traditional methods, our model significantly improves detection precision. Leveraging a weakly supervised Multiple Instance Learning (MIL) framework, we treat surveillance videos as collections of ranked clips, enabling efficient anomaly identification with minimal manual labeling. Optimized for real-world deployment, this scalable and high-performing solution sets new standards in public safety technology through intelligent, context-aware video analysis. Code.
PDF Code
Weakly-supervised Anomaly Detection in Surveillance Videos (MSc Thesis)