I built PH5, a multimodal phishing detection platform that combines NLP, URL classification, and computer vision. Its adaptive fusion engine achieved 99.5% URL and 97.2% text classification accuracy.
For PH5, I also implemented an OCR anti-evasion pipeline and optimized concurrent multi-model GPU inference, cutting memory usage by 65%. I served the models through FastAPI REST APIs.
I also built a RAG system for private-PDF question answering with inline citations, and AgentDesk CRM, where three LLM agents pause for human approval before database writes.
