Names, with a model

Add a model-based detector for people's names, wire your Aadhaar recogniser into Presidio the way a team would at work, and compare regex-only against hybrid on precision, recall and milliseconds. The artifacts are piiwall/names.py, a hybrid row in the results table with a PERSON number in it, and reports/latency.json.

Try first

Take ten messages from data/messages.jsonl that contain a PERSON span. Without a model, try to write a regex for names. Capitalised word pairs? Priya Nair yes, but also Video KYC, Reference ORD, Pay via. Words after Patient or Applicant? Works on the templates, fails on the first real chat.

Write down why names are a different problem from the five identifiers so far, in one sentence. Then write down what you expect an English NER model to score on Indian names in Hinglish text: pick a recall number. You will check it against reality at the checkpoint.