My MS capstone project at Indiana University South Bend.
The problem
Applications increasingly send user data straight to third-party AI services. That creates two risks: sensitive data (PII) leaking to the model provider, and the model being manipulated by malicious prompts or returning unsafe output.
What I built
A middleware service that sits between clients and AI services and inspects traffic in both directions:
- Inbound: detects PII and malicious or injection-style prompts before they reach the model.
- Outbound: screens model responses for unsafe output before they reach the client.
- Routing: securely routes and processes requests between clients and one or more AI services.
Results and lessons
Write-up in progress.