Consultant Data Scientist
Springer Nature Technology And Publishing Solutions · Nov 2024 – Present
- Deliver NLP and Generative AI solutions for publishing and pharmacovigilance workflows, covering data preparation, model design, retrieval and prompting pipelines, evaluation, deployment, and handoff to internal users.
- Prepared and normalized millions of publishing records, applied domain-aware document chunking, and generated dense embeddings.
- Indexed vectors with source metadata in Qdrant and implemented query embedding, metadata filtering, and similarity retrieval.
- Tested representative searches and refined chunk size, overlap, embedding choice, and retrieval settings before workflow integration.
- Defined Conference and Media Release label criteria, curated and labelled the dataset, and standardized the training format.
- Fine-tuned a Large Language Model (LLM), separated training and evaluation samples, and reviewed class-level errors.
- Refined ambiguous examples, training data, and model configuration through repeated evaluation cycles for consistent content routing.
- Ingested and cleaned internal API documentation, segmented it into retrieval-ready passages, and generated embeddings.
- Stored vectors with document metadata and implemented query retrieval, context assembly, and source-grounded prompt templates.
- Tested common developer questions and delivered the pipeline through an internal context-aware documentation chatbot.
- Mapped the source sections required for each summary and normalized complex, long-form drug-profile content.
- Developed prompt templates with explicit structure, scope, and output constraints for consistent summary generation.
- Reviewed outputs for completeness, factual alignment, and readability; incorporated feedback and deployed the workflow for pharmacological review.
- Parsed incoming pharmacovigilance XML files and mapped relevant elements for structured field extraction.
- Normalized safety content into reviewable records and passed structured data through a controlled narrative-generation workflow.
- Organized extracted fields and generated narratives for reviewer inspection while retaining access to the underlying source information.