Papers

This repository is a living ledger of my hands-on implementations of influential papers in AI, machine learning, and deep learning. It is a place to turn ideas from research into working code, experiments, and small, inspectable systems.

The collection will grow across areas such as Machine Learning, Deep Learning, new architectures, NLP, multimodal learning, inference, optimizations, AI safety, evaluations, and mechanistic interpretability. Each implementation lives in its own directory and documents what was actually built; an entry may reproduce a paper faithfully, explore one of its ideas, or extend it with experiments. Results are not assumed to be benchmark reproductions unless the implementation explicitly says so.

Paper implementation ledger

Paper	Implementation	What was implemented
Attention Is All You Need — Vaswani et al. (2017)	transformers	A from-scratch PyTorch encoder–decoder Transformer for English→Italian machine translation. It includes learned token embeddings, sinusoidal positional encodings, multi-head scaled dot-product attention, encoder/decoder blocks with masking and residual connections, training on Helsinki-NLP/opus_books, checkpointing and TensorBoard metrics, greedy autoregressive decoding, and notebooks for inference and attention visualisation.

Adding an implementation

To add a paper to the ledger:

1. Create a self-contained directory using a clear, lowercase slug (for example, paper-name/).
2. Include a directory-level README.md with the paper citation, setup and run instructions, and an honest description of the implemented scope.
3. Add one row to the table above, linking the original paper in the first column and the implementation directory in the second.
4. Describe only the parts that are present in code, experiments, or notebooks so the ledger remains a reliable record of the work.

Use this row as a template for future entries:

Paper	Implementation	What was implemented
Paper title — Author et al. (Year)	directory-name	Concise description of the model, experiment, tools, and analyses that were actually implemented.