{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreihrwazx5fhuto4vnzbsxkqzmwrmqki5qtozozjzdbtvrosirpuycy",
    "uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mkkjv5dkfns2"
  },
  "path": "/t/what-we-learned-building-a-privacy-first-layer-for-llms/175627#post_1",
  "publishedAt": "2026-04-28T11:38:30.000Z",
  "site": "https://discuss.huggingface.co",
  "textContent": "Hi everyone\n\nAfter experimenting with PII anonymization pipelines, we started building a more structured approach to using LLMs with sensitive data.\n\nA few things that surprised us:\n\n  * Naive regex + NER breaks quickly at scale\n\n  * Context loss can hurt model outputs more than expected\n\n  * Re-identification pipelines get tricky in multi-step workflows\n\n\n\n\nWe ended up moving toward a design where:\n\n  * sensitive data is abstracted before inference\n\n  * mappings are handled separately\n\n  * models never see raw PII\n\n\n\n\nCurious how others are approaching this—especially in production settings.",
  "title": "What we learned building a privacy-first layer for LLMs"
}