foundational machine learning research
Almaty, Kazakhstan · ra312.github.io · akylzhanov.r@gmail.com
Mathematician and research engineer working at the interface of ** harmonic analysis** and machine learning.
Core question: How do neural representations transition between monosemantic and polysemantic regimes, and what structures govern this behaviour?
My primary focus is mechanistic interpretability — understanding how internal representations evolve during training, how monosemantic features emerge and break down, and how to build a more rigorous grounded theory of neural feature representation structure rather than solely relying on empirical heuristics alone.
I use log-signature representations of sequences as structured probes of how transformers compress contextual information. Path signatures provide a graded algebraic basis that isolates contributions from different interaction orders.
I am interested to explore toy principles of feature superposition via lens of n-interacting particles system. The latter approach hasn been first developed here.
KazByte arXiv:2603.27859
Byte-level transformer adapter for Kazakh built on frozen Qwen2.5-7B.
Entropy-based dynamic patching (BLT-style) addresses the tokenizer fertility problem for agglutinative morphology. ~8M trainable parameters, trained on DGX-2 in collaboration with ISSAI, Nazarbayev University.
A zero-shot tabular transformer model Per-feature transformer pretrained on synthetic causal priors across 10+ financial institution datasets in Kazakhstan, UAE, Armenia, and Indonesia. Zero-shot tabular inference without task-specific fine-tuning. Trained across 8× RTX 4090 with FSDP2 + LoRA.
LogicNet Agentic tool-calling system with LangGraph orchestration for multi-step reasoning. Routes queries across fine-tuned models, symbolic solvers (SymPy), and API fallback based on topic classification and confidence thresholds.
12 peer-reviewed articles · 2 preprints · 220+ citations · h-index 7
-
Hörmander–Mikhlin type theorem on non-commutative spaces (2018) R. Akylzhanov, M. Ruzhansky, K. Tulenov · arXiv:2503.01240
-
Lp–Lq multipliers on locally compact groups (2020) R. Akylzhanov, M. Ruzhansky · Journal of Functional Analysis 278(3)
-
Smooth dense subalgebras and Fourier multipliers on compact quantum groups (2018) R. Akylzhanov, S. Majid, M. Ruzhansky · Communications in Mathematical Physics 362(3)
-
Hardy–Littlewood, Hausdorff–Young–Paley inequalities and Lp–Lq Fourier multipliers on compact homogeneous manifolds (2019) R. Akylzhanov, M. Ruzhansky, E. Nursultanov · JMAA 479(2)
→ Full list at Google Scholar
| 2024 – present | Senior Research Engineer, HighSky (highsky.io) — tabular foundation models, synthetic causal pre-training, LLM evaluation, agentic reasoning |
| 2022 – 2024 | Senior ML Engineer, Delivery Hero SE — transformer-based retrieval, RAG over 10M+ product catalogues |
| 2020 – 2022 | ML Engineer, KCell JSC — sequential modelling for 15M+ users, churn prediction, LTV estimation |
| 2018 – 2020 | Postdoctoral Research Associate, Queen Mary University of London (EPSRC EP/R003025/1) |
| 2017 – 2018 | Research Associate, Imperial College London (EPSRC EP/R003025/1) |
| 2014 – 2019 | PhD Pure Mathematics, Imperial College London · supervisor: Michael Ruzhansky · examiners: Fulvio Ricci, Ari Laptev |
| 2012 – 2014 | MSc Mathematics, Eurasian National University |
| 2007 – 2012 | Specialist Mathematics & CS (With Honours), Lomonosov Moscow State University |
- Kazakhstan Ministry of Science Grant (PI, under review 2026–2028) · £175K · Fourier multipliers on noncommutative spaces
- EPSRC EP/R003025/1 (Co-I, 2017–2020) · £500K · Harmonic analysis in semi-finite von Neumann algebras
- Doris Chen Merit Award, Imperial College London (2016)
- EPSRC Doctoral Studentship, Imperial College London (2014–2018)
Michael Ruzhansky (Imperial / Ghent) · Shahn Majid (QMUL) · Terry Lyons FRS (Oxford) · Yulia Kuznetsova (Franche-Comté) · Fedor Sukochev (UNSW, planned)
ra312.github.io · akylzhanov.r@gmail.com · +7 701 211 4844