Computer Engineering at Ain Shams University, Cairo (2027). I build and measure LLM systems: agents, tool calling, retrieval, and the evaluations that say whether they work.
Now
- Root-cause analysis for 5G O-RAN — graduation project with Cyshield: an LLM assistant that proves its diagnosis on an OCUDU testbed, or says it can't.
- AminBench — a tool-calling benchmark built from real Egyptian requests: Egyptian Arabic, Arabizi, and code-switched Arabic/English.
Selected work
- driftmatrix — restoring an LLM prompt cache changed an agent's action on 2–12% of ambiguous requests. Traced it to micro-batch tiling in llama.cpp; aligning the reuse boundary makes restores bit-exact on hybrid and dense models. Pre-registered, 6,480 server-verified runs.
- Chesterton — a mutation-testing reviewer for AI-written pull requests. It wrote execution-verified regression tests for 91 of 292 SWE-bench-passing patches; 80% also hold on the reference fix. Found an untested boundary in seaborn (PR #3999).
- Nephrolex — clinical guideline RAG inside the EHR (HL7 CDS Hooks): 0.71 nDCG@10, 0.70 on a held-out set, with bounding-box citations for every claim.
- UAV tracking, AIC-4 — team lead; 5th at the international final among 200+ teams, adapting a frozen transformer tracker at inference time only.
Python · PyTorch · llama.cpp · RAG · LLM evaluation · Docker