Siamak Barzegar

Built a next-generation recommendation system for automotive inventory, improving recommendation quality by 10–28% over the legacy approach. Cut model size by 75× with no loss in quality, reducing training and infrastructure costs. Developed a rigorous evaluation framework enabling evidence-based decisions across team ML experiments. Designed an intent-aware semantic search architecture allowing customers to search using natural language and retrieve vehicles matching their underlying intent. Modernised the ML training pipeline to improve reproducibility and speed up experimentation.

Architected REG-RA GENT, a proof-of-concept agentic RAG platform for natural-language analysis of large-scale regulatory knowledge bases, validating multi-step reasoning for cross-jurisdiction comparison, obligation analysis, and policy impact assessment. Implemented layered architecture using LangChain-based agent orchestration, hybrid retrieval pipelines, vector-based knowledge stores, and guardrail-based response validation to improve contextual accuracy and interpretability. Improved date-extraction KPI accuracy by 25% by leading migration from legacy services to fine-tuned LLMs (LLaMA 3, GPT, Gemini) with custom post-generation validation to eliminate hallucinations. Designed and shipped “RegGroup Summarizer” providing two-tier abstractive summaries with citation referencing and theme-based synthesis; engineered summarisation for hierarchical documents preserving parent–child section integrity. Built an LLM-based linking engine connecting regulatory changes to customer-speci

Built a clinical de-identification tool anonymising protected health information in GP and hospital reports, enabling access to sensitive datasets for research use. Developed automated MRI-report annotation and NLP-based “reason for visit” analysis supporting an integrated GP–paediatrician primary-care research programme. Delivered clinical-trial success-probability scoring to inform trial planning. Consulted on data-quality assessment for research programmes using NLP and deep-learning to surface insights and report impact to support programme delivery.