NLP/ML Engineer

Early-stage AI startup building an automated classification and risk assessment engine for a complex, multi-regime regulatory environment. The system ingests unstructured product and transaction data, extracts technical attributes, and classifies items against large government-maintained taxonomies with quantitative thresholds. The regulatory domain involves multiple overlapping classification systems, each maintained by different government agencies, with significant penalties for misclassification. Our system runs classifications across these taxonomies, correlates the outputs, and produces structured risk assessments designed for professional review. The core product. Specifically: • Information extraction pipeline that pulls structured, classifiable attributes from unstructured text (product descriptions, technical specifications, commercial documents) • Classification models that match extracted features against large hierarchical code systems (tens of thousands of entries across multiple overlapping taxonomies) • RAG architecture over a corpus of 500,000+ government rulings and regulatory documents • Vector similarity search for classification candidate retrieval and ranking • Technical parameter extraction and threshold matching against regulatory schedules (e.g., determining whether a quantitative product attribute crosses a controlled threshold) • Confidence scoring and calibration — the system must know when it’s uncertain and route to human review What We Need You to Know Required • NLP beyond prompt engineering — information extraction, named entity recognition, relation extraction, domain-specific fine-tuning • RAG architecture design and implementation (retrieval-augmented generation over large document corpora) • Embedding models and vector similarity search (FAISS or similar) • Python (this is a Python stack) • Working with technical and regulatory language, not casual text — precision matters more than fluency Strong Plus • Prior work in a regulated domain (legal tech, medical NLP, financial document processing, government data systems) • Experience with semi-structured document parsing (PDFs, tables, nested hierarchical schedules) • Experience building classification systems where accuracy has real-world consequences (not just recommendations) • Familiarity with evaluation frameworks for classification systems (precision/recall tradeoffs, confidence calibration, error analysis) • Comfort with ambiguity — we’re pre-launch, the architecture is designed but not all decisions are final What This Is NOT • Not a chatbot role. The system produces structured classification outputs, not conversational responses. • Not a wrapper around GPT. We are building domain-specific extraction and classification pipelines, not prompt chains. • Not prompt engineering. This is precision information extraction where a wrong classification has six-figure penalty consequences for our customers. • Not a research role. We need someone who ships production systems, not someone who writes papers about them.