online · Tokyo, Japan · UTC+9

Technical Profile

DennySORA

李汶道 DennySORA ~/dennysora/detail/depth

Technical Depth

How deep the knowledge goes: LLM architecture, training and post-training, inference and quantization, agentic AI, backend and distributed systems, cloud platforms and security.

← Back to Detail ↖ Resume

Technical Depth

LLM Architecture

Core

  • Transformer architecture
  • Decoder-only language models

Attention

  • MHA · GQA · MQA
  • latent-attention variants

Position Representation

  • RoPE · NoPE
  • hybrid positional designs

Feed-forward

  • SwiGLU · GEGLU

Normalization

  • RMSNorm

Residual Design

  • standard residual paths
  • experimental attention-residual variants

Tokenizer

  • SentencePiece · Unigram
Training & Post-training

Pre-training

  • token / data pipeline
  • optimization
  • learning-rate scheduling
  • checkpointing
  • validation loss / perplexity
  • throughput monitoring

Post-training · implemented

  • SFT · DPO · GRPO

Studied / compared

  • PPO · TRPO · DAPO
Inference & Quantization

Precision

  • FP16 · BF16 · FP8 · INT8 · INT4

Quantization Concepts

  • W / A / KV precision
  • PTQ · E4M3
  • W8A16KV8-style configurations

Runtime / Serving

  • vLLM · TensorRT-LLM · Ollama

Performance

  • KV cache
  • batching
  • context length
  • VRAM estimation
  • throughput / tokens per second

Exploring

  • MXFP8 · NVFP4
Model Interpretability CURRENT STUDY
  • Sparse Autoencoders (SAE)
  • SVD
  • activation analysis
  • activation patching
  • circuit-level interpretation concepts
Agentic AI PRODUCTION
  • Google ADK
  • Agent-to-Agent (A2A)
  • LangGraph · CopilotKit
  • AG-UI integration
  • tool execution
  • state management
  • Human-in-the-Loop
  • streaming (SSE)
  • retries & error handling
  • dynamic agent configuration
Backend & Distributed Systems

Languages

  • Go · Python · Rust

Backend

  • REST · FastAPI · Gin · Asyncio

Distributed Systems

  • Kafka · Redis · PostgreSQL
  • event-driven systems
  • asynchronous processing

Design

  • modular architecture
  • plugin architecture
  • microservices
  • low-latency systems
Cloud & Platform

Infrastructure

  • AWS · Kubernetes / EKS
  • Terraform · Helm

CI/CD

  • GitHub Actions · GitLab CI

Observability

  • OpenTelemetry · Prometheus · Grafana · Loki · Tempo

Production Experience

  • EKS node-group architecture
  • ARM migration
  • CPU / memory optimized pools
  • CNI troubleshooting
  • multi-tenant cluster stability
  • reusable Helm modules
Security
  • PII protection & redaction
  • policy enforcement
  • tenant isolation
  • access control
  • auditability
  • security architecture

Engineering Principles

01

Understand the Abstraction Below

Abstractions are useful, but debugging and optimization require understanding the mechanism underneath them.

LLM API
  ↓
Inference Runtime
  ↓
KV Cache
  ↓
Attention / GEMM
  ↓
Memory / Precision
  ↓
GPU
02

Deployment Is Part of Model Design

Architecture decisions affect memory, latency, throughput and deployment constraints.

Architecture
    │
    ├── memory
    ├── compute
    ├── KV cache
    ├── quantization
    └── serving
03

Measure Before Optimizing

Optimization should start from observable bottlenecks, not assumptions.

latency
throughput
memory
GPU utilization
cost
failure rate
04

Security Is Architectural

Security should be part of system design rather than a middleware added at the end.

Security
├── identity
├── authorization
├── isolation
├── policy
├── data handling
└── observability
05

Make Trade-offs Explicit

There is rarely a universally best architecture. The important part is knowing which constraints a design is optimizing for.

performance
    ↕
cost
    ↕
complexity
    ↕
reliability
    ↕
maintainability

How I Work

$ dennysora --working-style

research_first       true
prototype_fast       true
benchmark_driven     true
production_aware     true
security_conscious   true

default_loop:

understand → implement → measure → refine → repeat