I build LLM systems that are allowed to run in production, and the security tools that try to break them.

Two of mine already do: a pair of bots on a client's server that check and post job listings across Kuwait, Bahrain, Qatar and Oman every day. Getting a model to do that without quietly inventing things is most of the real work, and it's the part I care about.

Most of what's here started the same way, from a tool I wanted that didn't exist yet. ROPForge writes its own ROP exploit chains in Rust and clears ROP Emporium 8 for 8. My invoice pipeline hands a document to a human before it will auto-approve a total it can't verify. The safety gateway drops prompt injection and PII before they reach the model. I spend more time on evals, guardrails and failure modes than on the happy path, because that is the difference between a demo and something you can put real traffic through.

Final-year Electrical Engineering, NUST SEECS. Open to remote AI engineering and security roles.

📍 Islamabad, Pakistan  ·  📫 alimehdijafeeri@proton.me  ·  LinkedIn

• invoice-processing-system — vision-LLM invoice extraction with arithmetic validation and human-in-the-loop triage. Anything that fails validation is never auto-accepted.
• sales-research-agent — a ReAct agent that researches a company and returns a source-cited brief, with every claim traceable.
• codebase-intelligence-copilot — RAG over a code repository that answers in plain English and cites the exact file and line (recall@5 = 0.92).
• llm-eval-bench — an evaluation harness that gates CI, so a quality regression fails the build.
• ai-prod-monitor — online LLM observability: traces, token cost, a sampled LLM judge, and drift alerts.

• ai-safety-gateway — a drop-in proxy that puts prompt-injection, PII, and output guardrails in front of any LLM.
• prompt-injection-detector · jailbreak-eval-suite · persona-adoption-attacks · llm-guardrail-fuzzer · Crescendo-attack-detector — a toolkit for measuring and breaking LLM guardrails.

• ropforge — a Rust tool that automatically assembles working ROP exploit chains for x86-64 ELF. ROP Emporium 8/8, roughly 40x faster than angrop. Research paper in progress.
• wannacry-malware-analysis · redline-stealer-analysis · plugx-korplug-analysis — full static, dynamic, and reverse-engineering write-ups of real malware.
• network-traffic-analyzer · syn-port-scanner · traceroute-tool · ping-sweep-os-fingerprint · arp-spoof-detector · steganography-tool — a from-scratch network and offensive-security toolkit.

• nimbus-sdvn-fog — a UAV-hosted SDN controller for mmWave vehicular networks in fog (conference paper).

Tech: Python · Rust · C / C++ · TypeScript · LLM apps (RAG, agents, evals, guardrails) · Claude & OpenAI APIs · scapy · Ghidra · Linux