I have been writing code for fourteen years. No computer-science degree, no bootcamp, no mentor — just a lot of broken builds, a lot of reading other people's source code, and the stubborn habit of not moving on until I understand why something works.
I am not a researcher. I am a programmer who likes to know what is underneath.
My first programs were bad, and I did not know they were bad. For years I measured progress by "it runs" — which, I have since learned, is the least interesting question you can ask about software.
The turning point was not a language or a framework. It was learning to read the output instead of trusting that the program had done what I meant. Then learning to measure instead of guessing. Then learning that "it works on my machine" is not a claim, it is an apology.
Fourteen years later I still get things wrong. The difference is that now I find out sooner, and I write down what I found.
- Read the output. A green check mark is not evidence; the printed number is.
- Measure before you optimise. My best guess about where time goes is wrong surprisingly often.
- Fewer dependencies is a feature. Every one you add is code you did not write, cannot fix, and will still be debugging in five years.
- If you cannot explain it, you do not understand it. This is why I write things down — for the reader, but mostly to find out whether I actually know.
- The honest number is worth more than the impressive one. A small true result survives scrutiny; a big vague one does not.
Low-level systems work in Rust and CUDA: performance, memory layout, and the unglamorous layers that everything else is built on. Lately that has meant machine learning — not by calling a framework, but by building the pieces a framework would have given me.
If I had to point at one thing and say this is what fourteen years added up to, it would be this.
A language model written entirely from scratch, with zero dependencies: hand-written autograd, a byte-level BPE tokenizer, a RoPE/RMSNorm/SwiGLU transformer, our own CUDA driver FFI, cuBLAS and NVRTC bindings — and a 222M-parameter model trained on 2.05 billion tokens on a single consumer GPU. It took five months.
$ cargo tree
stonefish_ai v1.0.0That is the whole dependency tree. No torch, no candle, no cudarc — the CUDA
driver is loaded at runtime through the operating system's own loader.
| Parameters | 222,331,904 |
| Training data | 2,047,457,313 tokens — TinyStories + Project Gutenberg, exactly one epoch |
| Dependencies | 0 |
| Tests | 75, including 32 numerical gradient checks |
| Quality | BPC 0.6064 on held-out data (random baseline 2.909) |
| PyTorch parity | same weights, same batches → loss differs by 1.8e-05 |
| Throughput | 10,772 tok/s at 222M on one RTX 5070 |
| Code | 10,328 lines of Rust, 711 lines of CUDA, 1,415 lines of FFI |
$ cargo run --release -- demo 90
# trains a small GPT from scratch and writes Shakespeare in 90 secondsIt is not a chat assistant, and the README says so in the first screen. It continues English text; it does not answer questions, and it states invented facts fluently. Being clear about that was part of the work.
→ Repository · → English README · → Turkish README
bf16mixed precision and gradient checkpointing — the honest path to a ~400M model on the same 12 GB card- A from-scratch Turkish language model, with a browser demo
- Reading more source code than I write. Still the fastest way to learn.
📍 İzmir, Turkey
I am not on social media. If you want to reach me, open an issue — I read them.