Machine Learning Library in C

The brief

One level below NumPy: what does a matrix actually cost? This library builds the ML stack from raw memory up — a matrix type over a flat double*, the operations an MLP needs, and the MLP itself, in C with no dependencies.

The build

• Flat-array matrix type — one contiguous malloc(rows * cols * sizeof(double)), indexed as data[i*cols + j]; no pointer-to-pointer rows, so the whole matrix is one cache-friendly block
• Core ops — mat_mul, transpose, element-wise map, broadcast add — each with explicit ownership rules about who frees what
• MLP on top — dense layers + sigmoid/ReLU, forward and backward passes calling the same matrix ops, trained on small classification sets
• Valgrind-clean — --leak-check=full reports zero bytes leaked across the full train/predict/free lifecycle; every constructor has a documented destructor

What I learned

Manual memory management makes API design honest — every function signature has to answer "who owns this?" before it compiles into anything trustworthy. And i*cols + j taught me more about why NumPy is fast (contiguity, stride math) than any blog post had.