Build a pure‑Python speculative decoding engine that uses a draft LLM and parallel acceptance‑reject sampling to speed up text generation. This project highlights real systems engineering skills for your CV.
Learn how to build a high-performance INT8 convolution kernel from scratch using strided memory access and SIMD instructions.
This guide walks you through implementing a GPTQ-style post-training quantizer entirely in NumPy. The resulting script compresses transformer weights and measures accuracy impact, showcasing your ability to ship low-level ML infrastructure.
Build a sparse Mixture‑of‑Experts inference engine from scratch in Python, mastering top‑k gating, capacity‑constrained dispatch, and load‑balancing loss.
Implement a Switch Transformer‑style sparse mixture‑of‑experts layer in PyTorch, complete with differentiable routing, capacity limits, and load‑balancing loss.