~/talk

A DP Framework for Gaining Insights into AI Chatbot Use

Talkat PEPR '26 · 2026-06-01..06-02· [privacy-and-compliance, machine-learning]

speakers

name:Mikhail Pravilov

github:RamSaw

name:Edith Cohen

links:none on file

name:Vadym Doroshenko

links:none on file

name:Badih Ghazi

links:none on file

name:Charlie Harrison

links:none on file

name:Peter Kairouz

links:none on file

name:Pritish Kamath

links:none on file

name:Alexander Knop

links:none on file

name:Ravi Kumar

links:none on file

name:Ethan Leeman

links:none on file

name:Daogao Liu

links:none on file

name:Pasin Manurangsi

links:none on file

name:Adam Sealfon

links:none on file

name:Da Yu

links:none on file

name:Chiyuan Zhang

links:none on file

resources

video:https://youtu.be/5xmpDRDPDXo

slides:none on file

summary

Urania converts millions of raw chatbot logs into DP-safe summaries and trends, claiming end-to-end ε,δ DP guarantees by privatizing cluster centers and keyword counts rather than raw inputs. Pipeline uses LLM prompts for facet extraction, Gecko embeddings for semantic vectors, a differentially private clustering step, DP partition selection and histogram summarization, and LLM-based cluster naming. It runs distributed on Apache Beam + TPUs to process millions of examples in under a day and matches useful insight quality against a non-private Clio-like baseline.

AI-written summary · about

chapters (8)

related talks