Software engineer in Karachi. I build AI agent security tools, backend systems, and firmware for nRF hardware.
These days most of my time goes into agent security and evaluation: scanning agent skills before they run, and measuring how well other scanners and models catch malicious ones.
Agent security and evaluation: agent-skill scanning, static analysis, LLM adjudication, benchmark harnesses, SARIF AI/ML: PyTorch, computer vision (YOLO, RF-DETR, ByteTrack) Backend: Python, Django REST, FastAPI, PostgreSQL, Redis, Docker Web: TypeScript, React, Next.js, Tailwind Embedded: C++, nRF54L15, BLE, Zigbee, FPGA edge inference (Vitis AI) Systems: Rust
clawvet (PyPI, npm) Security scanner for AI agent skills. Static triage first, LLM adjudication second. Two independent research papers used it as their baseline.
jev-skillbench Benchmark harness for TypeSafe's Jev model as a malicious-skill detector. Ran all 7,944 skills in MalSkillBench across three inference backends.
unix-ancillary (Rust) Safe file descriptor passing over Unix sockets (SCM_RIGHTS). Used by the systemd-nspawn plugin for Forgejo's CI runner.
overruled (PyPI, GitHub Action) Verdict auditor for AI SOC agents. Replays ground-truth cases and grades the rulings, with no LLM in the grading path.
hwcontract (PyPI) Temporal assertions over hardware traces. Runs as a pytest plugin, a GitHub Action, and an MCP server.
- Zed editor: merged PR #48870, credited in v0.224.0-pre
- Roboflow: merged an RF-DETR and ByteTrack sports tracking pipeline
- "Comparative Assessment of YOLO Nano Architectures for High-Speed Steel Detection"
- "Secure Edge Deployment of Machine Vision System on FPGA Platform", IEEE CW 2026, accepted
