I'm Raed, an AI & Data Science graduate interested in machine learning, language models and the engineering behind intelligent systems.
Currently exploring how AI systems are built, evaluated and improved.
A character-level GPT built from scratch in PyTorch, following the core ideas behind the GPT architecture. The model is assembled from PyTorch primitives, without a transformer library.
Reaches 1.77 training and 1.90 validation loss on Shakespeare text. An early loss of 0.06 was too good to be true: it came from three bugs, one of them a missing causal mask.
A local RAG pipeline that retrieves relevant document chunks and uses a language model to generate grounded responses. Built and tested locally, without paid API services.
Passed 8 of 8 labelled retrieval checks on a small, five-document test set. When the retrieved context doesn't answer a question, it says so instead of guessing.
An experiment comparing conventional fine-tuning with parameter-efficient LoRA fine-tuning of DistilBERT, on six-class emotion classification.
Full fine-tuning scored higher in this run. LoRA came within 1.5 points of its accuracy while training about 1.10% of the model's parameters.
| Metric | Baseline | LoRA |
|---|---|---|
| Accuracy | 93.15% | 91.65% |
| Macro F1 | 88.86% | 87.71% |
| Trainable parameters | 100% | 1.10% |
| Training time | ~91 s | ~66 s |
In progress
A lightweight evaluation framework for testing and measuring retrieval-based AI systems. It explores how AI systems can be evaluated systematically, rather than relying only on subjective output quality.
Next: final validation, a documentation review and cleanup for reproducibility.
I'm interested in the practical engineering of machine learning systems, and I learn by building things and checking whether they work.
So far that has meant writing a transformer from scratch, experimenting with parameter-efficient fine-tuning, building a retrieval pipeline and exploring how AI systems can be evaluated systematically.
I'm particularly interested in why models behave the way they do, where they fail and how to make them more reliable.
If you'd like to talk about a role, a project or anything here, I'd be glad to hear from you.