Smit Radadiya

I’m a Machine Learning Engineer focused on turning LLM/NLP capabilities into reliable, production-ready systems. At HyperNorm AI, I build end-to-end pipelines that convert noisy content into structured insights for market and portfolio decision-making.

I built a multi-stage news processing pipeline on Apache Airflow with embedding-based clustering, LLM summarization, RAG-based topic tagging, and multi-dimensional scoring for market event ranking. I also designed real-time topic deduplication using sentence embeddings and connected components, plus ontology topic splitting that uses embedding similarity and LLMs to dynamically expand coarse nodes into finer subtopics.

To strengthen downstream reasoning, I designed an LLM-powered equity-to-equity knowledge graph mapping inter-company relationships and bidirectional equity impacts from news co-mentions. I built agents and engines for document parsing and constraint validation—turning investor mandates into executable Python validation functions using a Maker-Checker pattern with sandbox execution and LLM-as-Judge—and an agentic SQL rule engine that translates natural-language filters into SQL with multi-stage validation via EXPLAIN checks and LLM-as-Judge.

I also work on productionization and scalable infrastructure, including SQS-driven pipelines for earnings call analysis using semantic chunking into Milvus for RAG retrieval and FastAPI microservices for ML pipeline endpoints. Alongside engineering, I bring a research mindset from my M.Tech work and hands-on projects like implementing a Transformer from scratch in PyTorch, building VAEs, training segmentation models with U-Net, and developing NLP + classical ML systems—plus I’ve served as a teaching assistant for Data Structures & Algorithms (DSA).