Ahmad Hudhud

I’m a Machine Learning Engineer focused on turning messy data into reliable model behavior—especially for agentic, RAG, and fine-tuning pipelines. In my recent roles, I’ve worked across SDR-focused LLM fine-tuning, sales-signal extraction, and infrastructure that keeps training and inference efficient.

I’ve fine-tuned 3 specialized LLMs using LoRA/PEFT, improving instruction-following by +40% and cutting hallucinations by -70%, while reducing cost versus AWS Bedrock. I also trained a BERT transformer classifier achieving 83% F1 for sales-signal extraction, and built fault-tolerant, multi-step data generation that labeled 50,000 context pairs in 4 hours.

I care about measurable quality and operational excellence: I cut fine-tuning compute costs 3x and reduced VRAM usage by 80GB by optimizing distributed training and migrating to spot instances. From AWS SageMaker inference benchmarking and quantization (BnB, GPTQ, AWQ) to human-in-the-loop evaluation scoring (9/10), I aim for systems that ship, scale, and stay dependable.