← ~/projects
$ cat projects/genai-dlp.md
GenAI Data-Loss Prevention System
Self-hosted DLP for GenAI workflows: stop PII and secrets before they reach an LLM.
Sole developerJuly 2026Security
// why: Teams paste sensitive data into chatbots with no Indian-PII guardrail.
$ cat projects/genai-dlp.md▌
TL;DR
- Stops leaks: secrets + Aadhaar/PAN/GSTIN caught before the LLM
- Enforces policy: BLOCK/REDACT/JUSTIFY/WARN/ALLOW with hard floor
- Self-hosted: gateway + extension + Docker Compose, audited logs
// what does it do?
SIH self-hosted DLP between users and LLM providers. Detects secrets, Indian PII (Aadhaar, PAN, GSTIN), proprietary code — enforces policy via gateway, admin console, browser extension.
// why this project?
Employees paste sensitive data into ChatGPT-style tools every day, and Indian organizations have no self-hosted guardrail that understands Indian PII formats.
// what broke / learned
False positives were the real enemy — CI harness measures FPR on every push.
// results
- Four FastAPI microservices (detection, LiteLLM gateway, ingestion, HTMX admin) plus a Manifest V3 extension that intercepts prompts in the browser. Policy engine with BLOCK / REDACT / JUSTIFY / WARN / ALLOW actions and a BLOCK floor that roles can't weaken, HMAC-stable tokenization for redaction, content-redacted audit logs. Deployed with Docker Compose, with CI running a false-positive-rate evaluation harness.
4 FastAPI microservicesDetects Aadhaar, PAN, GSTINCI harness measures false-positive rate
Tech stack
FastAPILiteLLMDocker ComposeHTMXMV3 ExtensionPostgreSQL