I build AI agents, test them, and make them hard to break.
I spent the last year writing the tasks and graders that three AI labs use to evaluate coding agents, and I built an agent security system that a US client deployed in its own systems.
Try to break my agent
You are a researcher in the NLP lab. The agent may only use NLP and public documents. Pick an attack or type your own, and watch where it gets stopped.
This box runs the real input rules from my agent security demo, ported to JavaScript and checked against the Python code on 86 test prompts. The LLM judge and the output checks run on the server in the full system and are not simulated here. Nothing you type leaves your browser.
Three projects, in depth
Securing a shared research agent
An AI assistant shared by four research labs that answers with citations and never leaks one lab's data to another.
Read the case studyBuilding the tests AI agents are graded on
For a year I wrote the tasks, environments and automatic graders that AI labs use to train and evaluate coding agents.
Read the case studySentinel: following a vehicle across city cameras
My final year project. Many traffic camera feeds in one dashboard, plain English search over the footage, and a pipeline that links the same vehicle across cameras.
Read the case studyWhere I've done it
- AI Evaluation Engineer (contract), three AI labsNov 2025 to Oct 2026
Built agent evaluation tasks for three labs under NDA: agentic datasets (team lead, 5 people, 200+ tasks), Terminal-Bench style shell tasks, and SWE-Bench style repository tasks. Automated our task QA with LLM judges.
- AI Engineer, SWAM TechnologiesJun 2025 to present · project based
Agent security system for a US client, deployed in their systems. Agentic invoice validator.
- Founder, visa appointment automation serviceJul 2024 to May 2025
Built and ran a service that secured appointments for more than 500 clients. Reverse engineered undocumented booking portal APIs from network traffic and kept the automation reliable with session handling, retries and monitoring.
- BS Artificial Intelligence, FAST-NUCES Islamabad2022 to 2026
- Agentic invoice validator. Reads PDF invoices with an LLM, validates and cleans the fields, rebuilds a clean invoice and emails it. Python, OpenAI API, Gradio.
Three habits
Before I trust an agent I want a test set, a baseline and a number. I have written hundreds of graders, and I know most of the ways they get fooled.
Documents, tool outputs and users can all carry instructions. I design agents so no single check is the only thing in the way.
Docker, queues, retries, logs, and a README a stranger can follow.
Tools I use for real: Python, PyTorch, LangChain, MCP, FastAPI, Docker, Redis, pytest, OpenCV, Playwright, Linux.
Let's talk.
I'm looking for an AI engineering role, full time or contract. The fastest way to reach me is email.