🤝 Top Collaborators
.png&w=640&q=75)
🤓 Latest Submissions
VoiceShield
VoiceShield is built around a simple question: how can we be sure an AI voice agent is safe before letting real people talk to it? In real life, conversations don’t always go as planned. People can be angry, confused, vulnerable, emotional, or simply unpredictable. Traditional testing usually focuses on expected scenarios, so it can easily miss these difficult situations. VoiceShield tackles this by creating AI-powered personas that act like different types of real callers and put voice agents through realistic conversations. During each test, these personas challenge the voice agent with different situations while VoiceShield analyzes the conversation as it happens. It looks for things that could go wrong—such as unsafe or inappropriate responses, policy violations, risky behavior, or poor handling of vulnerable users. This helps teams discover potential problems before the AI agent is exposed to real customers. We use Speechmatics for voice interaction and speech processing, DataBright to bring relevant EU policies and regulatory information into the evaluation process, and our Smith evaluation agent to analyze conversations and determine risks, violations, and severity. The results are then presented through VoiceShield as clear Pass, Flag, or Fail outcomes with actionable recommendations. Our goal is to make AI voice testing feel less like traditional QA and more like putting an AI agent through a realistic stress test before real customers do it. VoiceShield helps businesses find problems early, understand where their agents are vulnerable, and build voice AI that people can trust.
10 Aug 2026
AgentGuard
Agent Guard is a real-time safety checkpoint designed to evaluate AI-generated clinical recommendations before they are presented for human review. Rather than replacing clinicians or medical AI systems, it acts as an independent verification layer that analyzes recommendations in the context of patient information and configurable safety policies. The application consists of a FastAPI backend and a Streamlit frontend. Users can enter patient demographics, symptoms, medical conditions, medications, allergies, vital signs, laboratory values, and a clinical question. The system can generate and verify a recommendation or verify an existing recommendation through the same workflow. The verification pipeline evaluates recommendations using modular safety checks and policy-based rules. Based on the collected evidence, the decision engine returns either PASS or FLAG together with a risk level, confidence score, explanation, and detailed verification results. Every verification is stored in an audit log to improve transparency and traceability. Clinical policies are stored in YAML files and can be reloaded without changing application code, allowing safety rules to be updated independently from the software itself. Audit records can also be viewed through the API. The project is fully containerized using Docker and exposes REST APIs through FastAPI while providing an interactive user interface with Streamlit. The architecture was designed to be modular so additional verification modules, policies, or language models can be integrated with minimal changes. The current prototype demonstrates how an independent governance layer can increase transparency and provide structured safety checks for AI-assisted clinical decision support workflows.
13 Jul 2026
