In-depth engineering notes on Transformer KV-cache topologies, DeepSeek MLA, vLLM high-throughput serving, quantization, and production Agent runtimes.
Curated deep-dives into system architectures and practical engineering
Why do traditional Selenium and Playwright scripts constantly break against modern web frontends? We dissect the architecture of Browser-use, the breakout open-source project with over 100k GitHub stars: DOM tree pruning, Set-of-Mark (SoM) visual coordinate grounding, state machine resilience, and end-to-end production recipes.
Read Full Article →Frontier technical notes and engineering practices
From physical silicon foundations and kernel matrices to 1:1 PD disaggregation topologies, DSpark dynamic speculative decoding, SWA bounded replay, deepseek-recipe protocol layers, and $38/GPU/h unit economics: a complete architectural dissection of DeepSeek's production inference stack.
Read Full Article →Production InfiniBand cluster operations: MLNX_OFED driver stacks, OpenSM high availability, ibdiagnet fabric health auditing, and NCCL GPUDirect RDMA tuning.
Read Full Article →AI cluster fabric topologies: Non-blocking Fat-Trees, Rail-Optimized cabling, Adaptive Routing (AR), SHARP in-network aggregation, and congestion mitigation.
Read Full Article →InfiniBand hardware architecture: SDR to NDR/XDR link speeds, 4b/5b to PAM4 modulation, credit-based flow control, Virtual Lanes, and OpenSM subnet management.
Read Full Article →Linux kernel networking internals: NIC RX Ring Buffers, DMA transfers, NAPI hybrid polling, sk_buff lifecycle, RSS multi-queue affinity, and eBPF XDP wire filtering.
Read Full Article →RDMA high-performance networking: Kernel bypass, hardware zero-copy, Queue Pair ibverbs verbs, and lossless RoCEv2 with 802.1Qbb PFC and DCQCN congestion control.
Read Full Article →Explore in-depth engineering breakdowns across the full LLM deployment lifecycle
DeepSeek MLA cache compression, MHA/GQA evolutions, RoPE encodings, and ultra-long context window limits.
vLLM production serving, PagedAttention memory virtualization, Chunked Prefill, and FP8/AWQ quantization.
Agent Runloops, Model Context Protocol (MCP), Skills mechanisms, sandboxed executions, and self-reflection loops.
Interactive 6-dimensional model radar, LMSYS Chatbot Arena, SWE-bench Verified coding benchmark, and ROI curves.
Engineering profile, technical creed, and research vectors
LLM Systems & Autonomous Agent Architect
This is my technical dossier documenting deep-dives into LLM foundation architectures, distributed serving post-mortems, and production agentic runtimes. Graduated with a B.S. in Computer Science from Beijing University of Posts and Telecommunications (BUPT), focusing on trillion-parameter inference serving, KV Cache memory topology modeling, and production-grade autonomous agent loops.
“Code must be verifiable, data must have provenance, and topology must be backed by mathematics. Moving beyond shallow prompts into rock-solid AI systems engineering.”