Ansar Zeinulla
Research & engineering · efficient ML · exact game solving
BSc Computer Science, Nazarbayev University (1st of 221) · Astana, Kazakhstan
I work on both sides of machine learning. In research, I solve games exactly and use the exact answer to measure how neural networks fail. In engineering, I write low-precision inference kernels that go into open-source software used in production. Every claim below has a number behind it and a public repository you can check.
News
- My llama.cpp pull request was approved by the reviewer from Arm.
- Interviewed on Astana TV.
- Talk at CMAI 2026, Nazarbayev University: Solving Bestemshe and Mapping Togyzkumalak to Benchmark Neural Decay.
- Revised Togyzkumalak paper submitted to the ICGA Journal, second round of review.
- The full solution of Bestemshe published: 8.96 GB dataset, any position one request away.
Stack
What I can do
- Train & fine-tune
- LoRA continued pretraining and supervised fine-tuning on one machine. Leakage checks and held-out splits — a number I cannot defend is worth nothing.
- Make it fast
- Quantised inference kernels in C++ with ARM matrix instructions. Before and after measured on an idle machine, wins and losses both reported.
- Data at scale
- Pipelines over 1011-row datasets, block compression, index design, records served over HTTP with no backend.
- Ship it
- CI, tests, browser demos, public model and dataset releases.
Research
First strong solution of Bestemshe
C++ · Python · ZSTD
A five-pit Kazakh mancala, solved exactly for all 125,698,242,527 positions: the second player wins. No one had solved it before. For scale, the strong solution of Awari (2002) took 778 GiB on a 72-node cluster; this one is 8.96 GB and answers any position over a single HTTP range request, with no backend at all.
LLM for a low-resource language, trained on a laptop
MLX · LoRA · Hugging Face
A 2.35M-word corpus for Nogai — 86,000 speakers, in no major model — built by a reproducible pipeline, then Qwen2.5-1.5B adapted to it on Apple Silicon. No cluster. I found train/validation leakage in my own data, rebuilt the splits, and withdrew a journal submission rather than publish numbers I could not defend.
Publications & talks
Combinatorial State-Space and Empirical Game Tree Complexity of Togyzkumalak: A Billion-Game Analysis — abstract in the Book of Abstracts of CMAI 2026, the 1st International Conference and Workshop on Mathematics and AI, hosted by ISSAI at Nazarbayev University, pp. 51–53, ISBN 978-601-08-6470-2 (PDF). Presented as the talk Solving Bestemshe and Mapping Togyzkumalak to Benchmark Neural Decay, 9 September 2026.
Journal article on the state-space complexity of Togyzkumalak — under review at the ICGA Journal (second round).
Engineering & open source
Faster ternary LLM inference for llama.cpp
C++ · ARM i8mm · SIMD
An 8×8 blocked matrix-multiply kernel for the TQ2_0 ternary format, replacing the scalar path. Pull request to a project with 90k+ stars, approved by the Arm reviewer and awaiting merge. Along the way I found the 13% slowdown it caused when BLAS is present and fixed it in the model loader.
Apple's AMX coprocessor, from Python
C · Python
amxpy exposes Apple's undocumented matrix coprocessor to Python, benchmarked across shapes and published with the cases where it loses (0.70×) as well as where it wins (2.91×).
C++ search engine with a WebAssembly build
C++ · WASM · CI
Search with transposition tables at 2.42M nodes/s in the browser, full CI, deployed demo. A reviewer found a real bug in my transposition table; I measured its effect, fixed it, and added a regression test that fails against the old behaviour.
Education
BSc Computer Science, Nazarbayev University — 3rd year, graduating June 2028.
GPA 3.87 / 4.00, ranked 1 of 221. Full state scholarship. Former research assistant.
First graduate of Tomorrow School (alem.ai).
Beyond research
First-category chess player · founded the Tomorrow School chess club · trainer and arbiter intern, Central Chess Club "Astana" (2022)