Skip to content

credgate

Governance-first PII detection for LLM requests, with Indian identifiers built in.

credgate finds personal data in text and tables before it leaves the machine for a language model, and proves that it did. It is built for data scientists in regulated Indian environments - banking, insurance, fintech and healthcare - under the DPDP Act 2023 and RBI, SEBI and IRDAI expectations, including fully air-gapped setups.

It ships as an R package and a Python package (with a command line) that read one shared specification, so both languages find exactly the same matches, write the same vault, audit and inventory files, and build the same queries.

What it does
Detect 24 Indian identifiers - Aadhaar, PAN, GSTIN, cards, accounts, UPI, mobile numbers, voter ID, passport, driving licence, ABHA, CKYC, ration card, NREGA job card, names, addresses and more - with checksums and context, every match explained
Scan and inventory Which columns hold which identifiers, in data frames, Arrow, Parquet, DuckDB, databases, Spark and BigQuery, sampled where the data lives
Redact and tokenise Replace identifiers before text reaches a model; keyed tokens give the same token for the same customer everywhere; encrypted vaults restore them afterwards
Policies Per endpoint (local, internal, frontier), what may be sent, redacted or blocked - for example card number with CVV and expiry never reaching a frontier model
Audit log A tamper-evident record of every decision, with no values in it

Everything runs offline; credgate makes no network calls of its own.

Install

pip install credgate              # Python 3.11+

install.packages("credgate",
  repos = c("https://dwarpal.r-universe.dev", "https://cloud.r-project.org"))

With Nix: nix build github:dwarpal/credgate#credgate (command line and Python) or #credgate-r.

The source, issues and changelogs are on GitHub.