Full-Stack AI/ML Engineer for Clinical Coding

We're looking for an experienced ML/NLP engineer to improve the accuracy of our clinical coding pipeline. We currently use three Hugging Face models to extract medical entities and match ICD/CPT codes from clinical notes, but they're missing codes that are explicitly present in the record and sometimes matching incorrect or overly generic codes (e.g., returning 11004 instead of 11042, or missing explicit codes like L97.228). Current stack: • d4data/biomedical-ner-all – entity extraction (diseases, symptoms, procedures) • emilyalsentzer/Bio_ClinicalBERT – embeddings for ICD matching • yikuan8/Clinical-Longformer – embeddings for CPT matching on long documents Scope of work: 1. Add a code-extraction step to catch ICD/CPT codes already written in the record (regex + validation against our existing mapping tables) — quick win, high priority. 2. Evaluate and swap in stronger pre-trained clinical models (e.g., BioBERT, UMLSBert, PubMedBERT, or clinical-specific NER models) to replace the current three. 3. Improve CPT matching so it's context-aware rather than keyword-based (e.g., correctly distinguishing mechanical vs. necrotizing debridement). 4. (Optional, phase 2) Fine-tune models on a labeled dataset of ~500–1,000 of our medical records with correct ICD/CPT labels. What we need from you: • Strong experience with Hugging Face Transformers, PyTorch, and biomedical/clinical NLP models • Experience with NER, embeddings-based matching, and evaluation of model accuracy • Comfortable working with medical coding standards (ICD-10, CPT) — healthcare NLP background is a strong plus • Ability to work with our existing Python codebase (ModelLoader, mapping tables) and propose/implement improvements incrementally Deliverables: • Working code extraction module for explicit codes • Updated model integrations with measurable accuracy improvement on test records • Documentation of changes and evaluation results • (If scoped) fine-tuning pipeline and fine-tuned model weights Please share relevant experience with clinical/biomedical NLP models and, if possible, a brief note on how you'd prioritize the steps above (quick fix vs. model swap vs. fine-tuning).