About the role
Frame ambiguous R&D and regulatory questions as well-posed modeling problems, selecting the appropriate method — classification, sequence modeling, information extraction, retrieval, generation, or statistical inference — and justifying the choice against alternatives.
Build and evaluate NLP and LLM systems over large corpora of scientific and regulatory text: document parsing and layout-aware extraction, named entity and relation extraction, semantic search and reranking, retrieval-augmented generation, structured output, and multi-step agent orchestration with tool use.
Develop predictive and statistical models over structured scientific data — analytical method and stability data, formulation and process parameters, historical submission outcomes — including feature engineering, handling of small and imbalanced datasets, uncertainty quantification, and model interpretability.
Mine historical health authority correspondence (deficiency letters, information requests, complete response letters) to learn recurring deficiency patterns and build submission risk-scoring models.
Design and own the evaluation discipline for every system shipped: curated golden datasets, inter-annotator agreement, failure taxonomies, offline and online metrics, A/B and regression testing, and quantitative acceptance criteria agreed with the business.
Build reusable modeling components — embedding and retrieval pipelines, extraction schemas, evaluation harnesses, fine-tuning and prompt-optimization workflows — on Databricks (Unity Catalog, Delta Lake, MLflow, Mosaic AI) and cloud model services (Amazon Bedrock, SageMaker).
Operationalize models with monitoring for data drift, performance degradation, and cost, and own remediation and periodic retraining.
Produce audit-quality technical documentation — model cards, data lineage, validation evidence — supporting Amneal's GxP AI validation framework and 21 CFR Part 11 requirements, working with QA on inspection readiness.
Establish benefit baselines with R&D and Regulatory owners and measure realized cycle-time, accuracy, and capacity outcomes.
Present methods, results, limitations, and trade-offs to technical and business audiences; mentor junior scientists and engineers on modeling practice.
Education
MS or PhD in Computer Science, Machine Learning, Statistics, Applied Mathematics, Computational Linguistics, or a related quantitative field; PhD preferred (exception, Bachelors from Tier-1 colleges).
Required Experience
5+ years building and deploying machine learning and NLP models in production, with demonstrated ownership from problem framing through monitoring.
Deep applied knowledge of modern ML and NLP: transformer architectures, embeddings, fine-tuning and parameter-efficient methods, retrieval-augmented generation, structured extraction, and LLM/agent evaluation.
Strong statistical foundations: experimental design, hypothesis testing, uncertainty quantification, and modeling with small, noisy, or imbalanced datasets.
Expert-level Python and the scientific ML stack (PyTorch or similar, Hugging Face, scikit-learn, pandas/PySpark); strong SQL.
Hands-on experience building agentic systems: multi-step orchestration, tool and function calling, and human-in-the-loop evaluation.
Track record of designing rigorous evaluation frameworks and defending model quality with quantitative evidence.
Ability to work within documented, controlled processes and produce audit-quality technical documentation.
Preferred Experience
Document intelligence over long, technical, or scientific documents (OCR, layout parsing, table and figure extraction from legacy PDFs).
Predictive modeling over chemistry, analytical, or process data; cheminformatics or chemometrics toolchains.
Experience in a GxP-regulated or other validated environment; familiarity with 21 CFR Part 11, CSV, or GAMP 5 and with AI model validation or governance.
Exposure to pharmaceutical regulatory content — ANDA/NDA/DMF, eCTD structure, SPL labeling, health authority correspondence. Not required; domain is taught on the job.
Databricks (Unity Catalog, Delta Lake, MLflow, Mosaic AI) and Amazon Bedrock or SageMaker.
Publications, open-source contributions, or competition results demonstrating applied ML depth.