agent runtime · retrieval evaluation · LLM serving · Go backend · production reliability
┌──────────────────────────────────────────────────────────────────────┐
│ │
│ 020617 → 22D3EE → 8B5CF6 → EC4899 │
│ DARK CORE CYAN SIGNAL VIOLET ROUTER NEON OUTPUT │
│ │
│ Build AI-native systems with backend discipline. │
│ Make it observable. Make it measurable. Make it ship. │
│ │
└──────────────────────────────────────────────────────────────────────┘
I build LLM-powered tools and backend systems around one simple idea:
AI products should be measurable, observable, and boring enough to run in production.
| Zone | What I care about |
|---|---|
| Agent Systems | ReAct-style loops, tool use, MCP-style integration, bounded execution, failure paths |
| Retrieval Infrastructure | Embedding, BM25, hybrid search, reranking, context construction, retrieval metrics |
| LLM Serving | vLLM-style serving concepts, SSE streaming, latency awareness, prompt systems |
| Backend Infrastructure | Go services, RPC integration, async DAG pipelines, Redis/MySQL, graceful fallback |
Languages Go · Python · Java
AI / LLM PyTorch · LangChain · vLLM concepts · Prompt Engineering
Retrieval Embedding · BM25 · Hybrid Search · Rerank · FAISS-style vector search
Backend RPC · Async DAG · Redis · MySQL · RabbitMQ · Object Storage
Systems Linux · Docker · Observability · Caching · Graceful Degradation
Evaluation Recall@K · Precision@K · MRR · nDCG · P50/P99 Latency
graph LR
A[Documents / Logs / Code] --> B[Parsing & Chunking]
B --> C[Hybrid Retrieval]
C --> D[Rerank & Context Builder]
D --> E[LLM / Agent Runtime]
E --> F[Tool Calls]
E --> G[Streaming Response]
F --> H[Observable Backend]
G --> H
- Agent runtimes with explicit tools, bounded loops, and clear failure paths.
- RAG pipelines evaluated with retrieval metrics instead of vibes.
- AI backend services that treat latency, caching, idempotency, and degradation as first-class constraints.
- Developer tools that compress repetitive engineering workflows into reliable agentic loops.
| Project | Signal |
|---|---|
| leetmate | Terminal LeetCode coach that gives hints, not answers. |
| gocommit | AI-assisted Chinese commit message generator. |
| llm-systems-learning-notes | Notes on LLM systems, inference, CUDA, scheduling, RLHF/GRPO. |
| minGPT | Minimal PyTorch GPT implementation and learning project. |
[01] Building small, composable AI Agent runtimes
[02] Improving RAG evaluation beyond demo-level retrieval
[03] Exploring LLM serving and streaming interaction patterns
[04] Keeping backend systems simple, observable, and failure-aware
[05] Writing tools that make developers faster without hiding the machinery
agent_runtime:
preference: explicit tools, bounded loops, observable traces
avoid: unbounded autonomy, hidden state, magic orchestration
rag_pipeline:
preference: measured retrieval quality, explainable context construction
avoid: demo-only vector search, vibe-based evaluation
backend:
preference: simple interfaces, boring infrastructure, graceful degradation
avoid: clever abstractions before the failure modes are understoodLESS MAGIC. MORE TRACES. BETTER SYSTEMS.