SmartGuard is a firewall layer for LLM applications, designed to run without GPU infrastructure.
What it does
- Classifies both inputs and outputs to catch jailbreak attempts, PII extraction requests, toxic content, and prompt injection patterns
- Configurable per-category thresholds to tune sensitivity without retraining
- Real-time Streamlit dashboard showing verdicts, accuracy, recall, and false positive rates
Testing
Built a red-team suite of 45 curated prompts benchmarked against baseline keyword filters, with P95 inference latency tracked to keep the classifier production-realistic.