Axel McKenna

Building systems by understanding them

I’m Axel McKenna. I build software that connects low-level engineering with practical applications. DAMASCUS is my exploration of what it takes to run a language model: loading its weights, turning text into tokens, computing a transformer forward pass, and checking every stage against an independent reference.

The engine is written in C++20, with Metal as the planned GPU compute path on Apple M1. The source and session write-ups document the implementation, the mistakes, and the reasoning behind each step. Correctness comes before performance claims.

DAMASCUS: from model file to logits

The verified CPU path loads Llama 3.2 3B Instruct in GGUF format, implements byte-level BPE tokenization, dequantizes Q4_K and Q6_K weights, and runs the 28-layer transformer. Its forward pass reaches per-layer cosine similarity of 1.0000000 against a HuggingFace reference using the same quantized weights, with exact next-token agreement.

Metal kernels and an AMX prefill path remain work in progress. The intended performance improvement is a research target, not a measured result. My technical write-up on building DAMASCUS explains the architecture, validation, and benchmark plan in detail.

Liquorfy: making prices comparable

I also created Liquorfy, which brings together liquor prices from New Zealand retailers. The engineering challenge is turning inconsistent retailer data into useful comparisons: identifying the same product across stores, normalising quantities, and showing prices near the user.

Source code and engineering notes

Explore Axel McKenna on GitHub, the DAMASCUS repository, and the session write-ups. For the project overview, return to DAMASCUS.