Ahmad H

AI/NLP Research Engineer

Freelancing Career • Indépendant

Sep 2024 - Present • 1 yr 11 mos

Conducting Master's thesis research on URBench — a comprehensive LLM evaluation benchmark for Urdu reasoning. Evaluated 11 models including Qwen3, LLaMA-3.1-70B, DeepSeek, Gemma, and Urdu-specialized models across 5 datasets using zero-shot, few-shot, and chain-of-thought prompting. Built full evaluation pipeline using vLLM on a SLURM cluster. Proposed SDFR-UR — a novel similarity-based dynamic few-shot retrieval method that achieved +6.00pp improvement on GSM8K and +4.94pp on PIQA over CoT baseline. Conducted RAG and cross-lingual prompting experiments with documented failure analysis.