Architect Dossier // ifnodoraemon
> LLM Infrastructure, High-Throughput Serving & Autonomous Agent Systems Engineer
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.”
> Engineering Creed // Core Principles
01. Mathematical Rigor & Memory Modeling
No blind guessing. Deep dive into physical KV Cache page virtualization, low-rank latent compression, and GPU communication topologies to mathematically quantify memory footprints and bandwidth saturation.
02. 100% Bare-metal Cluster Verification
No toy demos. All insights stem from hands-on debugging on 4-node 32-GPU H100 bare-metal clusters, troubleshooting mixed parallel (TP/PP/EP) deadlocks, Triton kernel exceptions, and memory fragmentation.
03. Environment Scaling & Sandboxed Agents
Moving beyond pure text conversations to embrace the 2026 Environment Scaling paradigm: Model Context Protocol (MCP) tool buses, MCTS reflection self-correction, and zero-escape Linux container sandboxes.
> Core Engineering Vectors
⚡ High-Throughput Serving & Deployment
vLLM / SGLang production serving, PagedAttention memory virtualization, Chunked Prefill, FP8 dynamic quantization, and high-concurrency API gateways.
🧠 Attention Topologies & Memory Wall
DeepSeek MLA (Multi-head Latent Attention) low-rank projection mathematics, Kimi Delta Attention state progressions, and RoPE long-context extrapolation.
🤖 Agent Systems & Runtime Engineering
Model Context Protocol (MCP) unified bus, Agent Runloops, MCTS-based self-reflection, and secure sandboxed Docker/eBPF runtimes.
📊 Model Benchmarking & Selection
SWE-bench Verified real-world code repairs, Terminal-Bench 3.0 sandboxed execution, LMSYS Arena evals, and price-to-performance frontier analysis.
> Flagship Production Systems
End-to-end engineering track record from high-concurrency enterprise B2B backends to autonomous Agent OS
⚡ Heterogeneous MaaS Serving Platform (K8s + GPUStack)
MAAS INFRACompute Architect & Cluster Engineer
Engineered cross-chip LLM serving foundations via K8s and GPUStack, managing heterogeneous clusters across NVIDIA (A800/H800), Huawei Ascend (910B/310P), and Iluvatar PPU 810E with optimized weight caching, MindIE/vLLM engines, and multi-tenant resilient routing.
🤖 Navi - Governed Personal Agent OS
AGENT OSCreator & OS System Architect
A local-first personal AI assistant OS featuring declarative capability manifests, Bubblewrap sandboxing, durable Goal/Run loops, and WeChat/Telegram connectors.
🏢 Nowcoder Enterprise HR & AI Resume Scoring (hr.nowcoder.com)
ENTERPRISE / AICore Backend Engineering / ML Scoring System
Java / Spring Boot enterprise microservices powering B2B recruitment workflows; engineered a Python scikit-learn (sklearn) SVM resume evaluation system for precise automated candidate scoring.
⚡ Doraemon Code (Autonomous Dev Agent)
DEV AGENTSingle-Process Local Code Agent
Open-source Claude Code alternative for local autonomous programming with 39+ native tool integrations.
> Academic & Infrastructure Foundations
🎓 Academic Foundation
Beijing University of Posts and Telecommunications (BUPT)
Computer Science and Technology
Specialized in computer architecture, operating systems, distributed computing, and deep learning, building the rigorous engineering foundation required for low-level systems work.
🖥️ Hardware & Heterogeneous Cluster Portfolio
Hands-on experience tuning enterprise MaaS heterogeneous serving clusters built on Kubernetes (K8s) + GPUStack. Deep operational mastery spanning NVIDIA (A800 / H800 / 32-GPU H100 bare-metal clusters), Huawei Ascend (910B / 310P NPUs), and Iluvatar (PPU 810E) accelerators, covering unified virtualization, weight caching, cross-platform MindIE / vLLM serving, and distributed parallel pipelines.
> Tech Stack & Matrix
LLM Systems, Heterogeneous Compute & Serving Engines
HPC / Serving / KernelsAutonomous Agent Systems & Runtime Protocols
Agent OS / MCP / SandboxingPython Engineering, Data Science & Machine Learning
Python 3.13 / ML / Computer VisionJava Enterprise Backend & Distributed Architecture
Enterprise B2B / High ConcurrencyFrontend, Cross-Platform Desktop GUI & Qt Engineering
Frontend / Qt Desktop / Visualization> System Metrics
// All whitepapers and reproducible benchmarks are continuously iterated. Connect via the following interfaces: