MD Asjad

Research & ML foundation Data Scientist with deep research experience building PyTorch-based, end-to-end modular systems for decision-making under uncertainty and noisy logged data, with offline evaluation showing strong gains over linear baselines. I specialize in fine-tuning LLMs and Transformers, building production deep learning models in PyTorch and TensorFlow, and optimizing inference with TensorRT and vLLM. I worked on production AI for marketing and customer data at scale. Built a metadata-indexed, multi-collection RAG pipeline with hierarchical 2-phase retrieval — parallel table/rule discovery, then dependent schema fetch — and a low-latency Knowledge Stack assembly layer that cuts schema noise and retrieval overhead so we own retrieval, context, and generation independently of legacy tooling. On top of that foundation, researched and shipped automated conversational analytics: multi-turn query rewrite, business-rule–aware SQL generation via DSPy, and downstream insight/report workflows. Designed and built an AI journey builder that converts natural language into structured, multi-channel marketing workflows — triggers, branches, channels, and targeting — through a single LangGraph planner with tool loops, recommendations, and validation baked in. The core innovation is manifest-driven schema automation: when the product UI adds or changes touchpoints, agents, validators, and converters regenerate automatically, while MCP-governed tools provide safe, production-grade access to audience data. I also design and operate lakehouse ETL pipelines using PySpark and the Delta stack—incremental MERGE ingestion from OLTP to ADLS, partitioning and optimization, and downstream profiling and analytics—so agents and insights run on reliable, tenant-scoped data foundations. Apart from that contributing to CI/CD and collaboration with teams.