Nehad Shakoor

An offline medical assistant that answers from a WHO corpus with no internet connection, running a quantized Llama 3.2 3B through llama.cpp. Smarter chunking lifted retrieval precision from 15% to 85%. Final-year project, two semesters, written up as an IEEE paper.

A model that resists prompt injection at full precision doesn't always hold up once it's quantized to int4. This measures the difference — twelve conditions, the same attacks, run across different precision levels.

A deliberately vulnerable Flask app for working through the OWASP LLM Top 10 one case at a time. Comes with a jailbreak-prompt classifier (logistic regression, 88% accuracy) as a baseline to test detection against.

A security scanner for LLM agents — the kind that can call tools and run commands. It checks for tool-poisoning, prompt injection through retrieved content, and data leaking out through tool calls. Point it at a running agent and it reports what it could get through.

A tool that turns long technical talks into clean, readable articles — I built it to get IBM Technology videos into a format I could actually reread. Built with the Claude API, FastAPI, Next.js, and Docker.

A benchmark for testing how language models hold up against attacks in Urdu — a language 230 million people use that most safety research skips. Native-speaker-checked prompts paired with English translations, scored on how consistently models resist and refuse across both.