Marwan AbdElhameed

AI for Software Engineering

Can LLM ensembles catch what single models miss in software artifacts?

Incomplete requirements are a quiet, expensive failure mode: what a specification leaves unsaid rarely surfaces until it breaks something downstream. I mapped how the field defines, measures, and tools requirements completeness, then showed that ensembles of diverse LLMs catch and repair gaps single models miss. Next I want to try the same ensemble approach on other software artifacts.

• ASE 2025
• SEAA 2025

Compute-Optimal LLM Reasoning

When is more inference compute actually worth it?

Spending more compute at inference time can substitute for model scale, but the trade-off is rarely measured carefully. With Pavly Halim, I compared inference-scaling strategies against reasoning-trained approaches on math problem-solving under matched compute budgets. The question gets more relevant every year: for deployed systems, inference is where the money goes.

• arXiv 2024

AI-Powered Requirements Analysis Platform

• ▹Architected intelligent SRS analysis platform integrating 3 LLM APIs with advanced aggregation strategies, achieving 95% consensus accuracy.
• ▹Engineered document processing pipeline supporting PDF/DOCX/TXT with Mistral AI OCR integration.

Math Problem-Solving LLM System

• ▹Fine-Tuned Mistral-7b achieving 35% accuracy on GSM8K dataset.
• ▹Enhanced solution accuracy by 25% from base model through advanced inference scaling techniques.

Distributed Model Serving Platform

• ▹Built scalable REST API handling concurrent model inference requests, achieving 99% uptime.
• ▹Integrated Ray with TorchX for distributed computing, processing 10,000+ concurrent requests.

Multilingual RAG System

• ▹Architected multilingual RAG system with mBART-large-50 for cross-lingual content generation.
• ▹Improved content relevance by 40% through context-aware document fetching.