∞
σ
Σ
∂
∇
dW
λ
∫
Δ
μ
ρ
θ
Agentic AI/ML · Quantitative Finance
Systems for models, markets , and scale. I'm Zhanyl Abdybaeva , an ML infrastructure and quantitative finance engineer
working at a quantitative trading firm. My work spans GPU clusters, inference systems,
distributed training, and the schedulers that decide what runs when.
This site is where I publish long-form technical writing, shorter field notes, and
project pages for the systems I'm building.
GitHub
LinkedIn
Work with me
About
RSS
Role
Infrastructure engineering at a quantitative trading firm
Focus
GPU clusters · inference · Slurm at scale
Studying
MS CS — ML @ Georgia TechCQF
Writing
Deep dives & lab notesmost weeks
ls ./writing --featured January 2026 · blog · 7 min read
Structured sparsity aligned with GPU attention kernels: LServe’s unified serving stack for prefill and decode, and SampleAttention’s empirical patterns plus CRA as a runtime quality floor.
January 2026 · blog · 4 min read
A technical walkthrough of PagedAttention: the memory management innovation that makes vLLM practical for serving many concurrent LLM sessions without fragmenting GPU memory.
All deep dives →
ls ./projects slurm-rca-bench
Shipped A public incident-diagnosis benchmark for HPC schedulers — ten reproducible failure scenarios with measured ground truth and degenerate baselines.
slurm hpc benchmark root-cause-analysis
Source →
Epilog GPU Validator
Shipped Drain a node for a persistently faulty GPU — and never for a transient one.
slurm gpu nvidia hpc
Source →
cluster-sre-agent
In Build Multi-agent diagnosis over an explicit cluster dependency graph — five ablatable configurations, specified before results existed.
slurm llm-agents mcp multi-agent
Source →
Slinky GitOps
Shipped Slurm on Kubernetes from nothing in one command — and the auth-key rotation nobody wants to test in production.
slurm kubernetes slinky helm
Source →
k8s-gpu-scheduler-lab
Phase 1 shipped A controlled comparison of Kubernetes GPU schedulers on the same workload traces, built on kwok so it runs without GPUs.
kubernetes gpu scheduling kwok
Source →
IB Slurm Exporter
Shipped Correlate InfiniBand and RoCE counters with the Slurm job that owns them, so a slow collective can be traced to the fabric.
slurm infiniband rdma prometheus
Source →
GPU-Accelerated Volatility Surface Calibration
Design Phase CUDA-accelerated local and stochastic volatility surface calibration — Dupire's model, Heston, and hybrid approaches for real-time derivatives pricing.
quantitative-finance cuda gpu volatility
GPU Reaper
Shipped Find wasted GPU allocations on a Slurm cluster and escalate them through alert, drain, and cancel.
slurm hpc gpu nvidia
Source →
Slurm Scheduler Lab
Shipped Test Slurm priority and backfill policy against a job trace before it reaches a live controller.
slurm hpc scheduling backfill
Source →
vLLM vs TensorRT-LLM: Inference Throughput
First pass — measured, single GPU A measured throughput comparison of vLLM and TensorRT-LLM on Llama-3 70B, and why the faster engine is not automatically the right one.
vllm tensorrt-llm inference benchmark
slurm-mcp
Shipped A read-only MCP server exposing Slurm scheduler state to agents, with the allowlist enforced in code and progressive disclosure of the tool surface.
slurm mcp agents hpc
Source →
cluster-ops-skills
Shipped Production HPC and GPU cluster runbooks packaged as loadable Agent Skills, each one naming the wrong diagnosis it exists to prevent.
slurm hpc agent-skills runbooks
Source →
research-platform
Shipped Point-in-time data semantics for quantitative research — as-of queries, feature lineage, and leakage detection.
python quantitative-research point-in-time data-quality
Source →
