Diviqra Guard

LLM firewall for production AI applications.

Detects prompt injection, jailbreaks, PII leakage, and indirect injection attacks. Under 10ms on CPU. No GPU required. Works with any LLM.

Install

pip install diviqra-guard

Quick start

from diviqra_guard import Guard

guard = Guard(api_key="dg_dev_...")
result = guard.scan("Ignore all previous instructions")
# ScanResult(blocked=True, threat="prompt_injection", latency_ms=8)

Why Diviqra Guard

Architecture

Three-wall defence:

• Wall 0 — Pre-processing (<1ms): multilingual normalisation, entropy analysis
• Wall 1 — DistilBERT ONNX (<10ms): 223K sample training, pattern rules, PII
• Wall 2 — LLM Judge (~20% traffic): qwen3:1.7b, context coherence, semantic drift

Integrations

LangChain

from diviqra_guard.integrations.langchain import GuardCallback
from langchain_openai import ChatOpenAI

llm = ChatOpenAI(callbacks=[GuardCallback(api_key="dg_dev_...")])
response = llm.invoke("Hello")

OpenAI SDK

from diviqra_guard.integrations.openai import wrap_openai
from openai import OpenAI

client = wrap_openai(OpenAI(), api_key="dg_dev_...")
response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello"}]
)

FastAPI

from fastapi import FastAPI
from diviqra_guard.integrations.fastapi import GuardMiddleware

app = FastAPI()
app.add_middleware(GuardMiddleware, api_key="dg_dev_...")

Benchmark

What Guard detects

Self-host

git clone https://github.com/diviqra-builds/diviqra-guard
cd diviqra-guard
pip install -e ".[service]"
cp .env.example .env
psql -d yourdb -f migrations/0001_guard_events.sql
./start.sh

Pricing

Get a free API key at guard.diviqra.com

Training data

Fine-tuned on MIT-licensed datasets from Lakera AI and deepset. Model on HuggingFace: diviqra/distilbert-guard

License

Core scanner: MIT

SaaS console: guard.diviqra.com

Built in India. Works everywhere.