AI Firewall

Web applications need to identify malicious or policy-violating inputs before those inputs reach the services behind them. This project investigates how compact language models can perform that detection at the application boundary, where inference cost and latency matter alongside detection quality.

My work spans the full model pipeline:

• Data representation: Train domain-specific tokenizers that capture patterns in the inputs the firewall must inspect.
• Detection models: Train small, BERT-family language models for security classification tasks.
• Inference: Quantize the models and deploy them for low-latency detection.

The goal is a practical detection component that can run close to the application, with model design and deployment considered together.