Senior AI/NLP Engineer — Scientific Document Understanding and Evidence Validation - Contract to Hire

We are building a Python system that reads scientific research papers and converts their original findings into structured, evidence-backed records. We need an experienced AI/NLP engineer to improve extraction quality, evaluate alternative interpretations of research findings, and make the pipeline reliable and cost-efficient. The work requires careful handling of scientific evidence. Each extracted finding must preserve the paper's scope, experimental conditions, population or species, quantitative results and uncertainty. Every finding must link to exact supporting text. The system must distinguish original findings from background citations, speculation and unsupported conclusions. You will receive a starter repository, a documented output schema, sample PDFs and parsed source data, development fixtures, a local evaluation workflow, and scripts for reproducible comparisons. Model-based evaluation requires configured inference access. The local evaluator supports iteration; it is not a guarantee of production performance. Responsibilities - Improve scientific PDF ingestion, including tables, captions, multi-column text and source locations. - Build and optimize an LLM-assisted extraction pipeline producing structured JSON with findings, supporting evidence, experiments and source references. - Preserve numerical details, causal language, uncertainty and study limitations. - Develop an evidence-verification component that compares alternative statements, selects the supported option and provides exact quotations. - Compare approaches using frozen development references and unseen evaluation papers. - Investigate errors and improve accuracy without increasing unsupported statements. - Track model usage, inference cost, latency, batch completion and failures. - Deliver maintainable code, reproducible setup, experiment results and clear handoff documentation. Required experience - Strong Python engineering skills. - Practical experience with LLM-based information extraction and structured outputs. - Experience designing evaluation datasets and diagnosing model errors. - Familiarity with PDF/document parsing and evidence-grounded systems. - Ability to reason about scientific findings, study scope and numerical results. - Experience with API integrations, logging, retries, concurrency and deployment. Helpful experience - Scientific NLP, biomedical literature mining, systematic reviews or knowledge graphs. - DSPy, LangChain or comparable agent frameworks. - GROBID, layout-aware PDF extraction, OCR or table extraction. - Model routing, prompt optimization and inference-cost measurement. Expected deliverables - A reproducible baseline and a concise assessment of the main failure modes. - An improved extraction and review implementation. - Paired evaluation results on a frozen benchmark and separate unseen papers. - A cost/runtime report and complete batch success/failure accounting. - Source code, setup instructions, configuration examples and handoff notes. This is an hourly engagement with no weekly hours limit. Progress will be reviewed through completed deliverables, documented experiments and reproducible evaluation results. The initial work will establish a reproducible real-paper baseline and test promising improvements. Quality will be assessed using reproducible results and independently checked examples. Please provide a short update each working day covering completed work, experiment results, blockers and the next planned step. Record code revisions and evaluation settings so another engineer can reproduce your results. Model/API spending will have a separate agreed budget. Please include in your application - One relevant extraction, document-AI or scientific-NLP project you personally built. - Your proposed approach to preserving the scope of a finding and linking supporting evidence. - How you would evaluate improvements when perfect ground truth is unavailable. - Your experience with scientific PDFs and tables. - Your hourly rate, availability and proposed approach to establishing the real-paper baseline and testing improvements. Apply To This Job