Timestral after 561 trades: confidence meets trading costs
Nine simulated accounts, 561 executed trades, and a fragile FX winner: what my Timestral-209M experiments actually showed.
Read articleBlogs
Engineering notes.
Notes on fine-tuning, inference, model serving, agent systems, and ML infrastructure.
Back to portfolioPushing the limits of legacy hardware
The RX 560X journey: VRAM surprises, Vulkan kernels, exactness bugs, 316.60 aggregate tokens/s, and a merged upstream contribution.
Read articlePagedAttention in vLLM: scattered memory, exact attention
Memory fragmentation, virtual memory, KV blocks, prefix reuse, and the exact math behind paged attention.
Read articleInference tuning starts with a workload
How Inference God Mode turns model, hardware, API, and workload requirements into a measured deployment and optimization loop.
Read articleBenchmarking real repositories with AnyBench
How AnyBench turned recent Humanize and Boltons commits into verified coding tasks, and what a small model run can and cannot tell us.
Read articleFine-tuning starts before training
Why I turned practical LoRA, Unsloth, quantization, evaluation, and vLLM field notes into an agent skill for Claude Code, OpenCode, and Codex.
Read articleBuilding a small fruit-fly brain in the browser
How FlySim handles odor identity, sparse memory, learned avoidance, compass drift, and a fixed 200 Hz brain loop.
Read articleTechnical articles · Zilliz
Published with Zilliz.
My technical articles on LLM reliability and vector databases.
View my Zilliz author pageDecoding LLM Hallucinations: A Deep Dive into Language Model Errors
Why language models hallucinate, and how self-evaluation, reference checks, uncertainty, and consistency can help detect errors.
Read on ZillizSpring AI and Milvus: Using Milvus as a Spring AI Vector Store
A practical walkthrough of connecting Spring AI to Milvus for vector storage and similarity search.
Read on ZillizPost title
Post summary.