Skip to content
View zainbabar's full-sized avatar
🕸️
🕸️
7 followers · 8 following

Highlights

  • Pro

Block or report zainbabar

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
zainbabar/README.md

hey, i'm zain.

second-year mathematics student at the university of waterloo.

i'm interested in ml, deep learning, and ai, from the math underneath to the systems built on top. lately that's meant implementing things from scratch and training models to see what actually happens.

website ↗

stack

languages

Python C++ C SQL

ml & data

PyTorch torchvision Hugging Face NumPy pandas scikit-learn Matplotlib LangGraph LangChain Chroma Ollama FastAPI

tools

Docker Git Linux Jupyter CMake pytest

selected projects

gpt-2 inference engine

gpt-2 small (124m) from scratch in pytorch, checked against hugging face. kv caching reached 175 tokens/s on an nvidia gb10, a 10.5× speedup for a 992-token prompt + 32 generated tokens.

autograd engine

a scalar autograd engine and neural net library from scratch in c++20. numerical gradient checks, xor training, and 93.8% mnist test accuracy without ml libraries.

lecture rag

a study assistant over my linear algebra notes, with cited answers and local models. reranking raised test recall@1 from 80.9% to 92.6%, measured on labeled evidence in a frozen 40-question set.

pathshift

a cnn trained on pathology images, with corruption tests and label-free shift detection. mmd caught the clinical-center shift in all 100 trials at 200 images per batch, with a 5% false-alarm rate on clean batches.

contributions

my contribution skyline

linkedin · email · x

Pinned Loading

  1. gpt2-inference-engine gpt2-inference-engine Public

    gpt-2 small built from scratch in pytorch, verified against hugging face. kv cache: 175 tok/s on nvidia gb10 (10.5x), 42 tok/s on m3 pro cpu (18x); 992-token prompt + 32 new tokens.

    Python

  2. chess-llm-reasoning chess-llm-reasoning Public

    A 1.7B model trained on verified answers beats gpt-oss-120b on 1,000 unseen chess puzzles (63% vs 53%). Trained on the 120B model's explanations instead, it got worse, and every RL reward that chec…

    Python

  3. lecture-rag lecture-rag Public

    rag from scratch: dense + bm25 hybrid search, and a langgraph evidence check, all measured with recall@k

    Python

  4. autograd-engine autograd-engine Public

    reverse-mode autodiff engine and neural net in c++20, built from scratch and verified against numerical gradients.

    C++

  5. chambercrawler3000 chambercrawler3000 Public

    object-oriented c++ roguelike built around five core design patterns

    C++

  6. pathshift pathshift Public

    cnn built from scratch in pytorch, with a robustness eval and label-free MMD shift detection.

    Python