Part 1 of 4 · 1 min read
Overview
Agents read customer data, write it to memory, send it to tools and log all of it. Interviewers ask "how do you handle PII in an LLM application?" The strong answer lists the copies, minimises first, tokenises, redacts logs, sets retention, handles erasure of derived data, and states the limits of detection.
About the example, and what is real
Real: the detectors (email, phone, SSN-format, card with Luhn, IPv4, titled names), the span-level scoring, the vault, the minimiser, the store with TTL and lineage-based erasure, and the 26 test cases in code/mini-privacy. Assumed: the 20 texts and their labels are fictional and written by me (including one deliberately conservative label: a part number in SSN format counts as PII). Not real: no real personal data, no ML-based detector such as Presidio or an NER model, no database, vector store or provider logs; erasure is modelled on an in-memory store.