yashpawar6849/Aegis ? reverse-engineered prompt

Build me a local Python app called Aegis that acts like a safety layer for an LLM while it is generating text.

I want a backend that can wrap a Hugging Face style causal language model, watch its internal activations during inference, and steer or block outputs when the model seems deceptive, overly aggressive, or unsafe. It should also stream live metrics as the model runs, so I can see what is happening token by token.

Please include a simple web dashboard where I can pick a model, run a few preset safety checks like deception, sycophancy, or harmful coding requests, and send my own prompt to see the results in real time. Make it easy to run locally, with clear setup steps and a single command to start the server and open the UI.

If you need current library details or best practices, look up the latest docs online.