Stella Shiina
Paused
36%
中文
Security × Systems × AI

Stella Shiina.

Making complex systems fast, reliable, and secure.

I build high-performance services, network infrastructure, and security tools with Go and TypeScript, focusing on measurable performance, explicit security boundaries, and reproducible experiments.

View experience Get in touch
Continue exploring

Capabilities / Evidence over ratings

Capability is shown in systems, not scales.

From concurrent pipelines and memory layouts to authentication boundaries, network protocols, and reproducible experiments. Each group is grounded in engineering evidence.

01

Systems & Performance

Building measurable systems from memory layout and concurrent pipelines to real network communication.

Go

Used for HTTP services, concurrent alert pipelines, authentication, and consensus-network simulation.

  • Gin
  • Goroutine
  • Worker Pool
  • Go Testing
Performance

Focused on throughput, latency, GC pressure, contiguous memory layouts, and object reuse.

  • sync.Pool
  • Profiling
  • Memory Layout
Networked Systems

Experience with TCP P2P, Gossip, WireGuard, and service gateway boundaries.

  • TCP P2P
  • Gossip
  • WireGuard
  • Caddy
02

Security & Infrastructure

Treating authentication, authorization, rate limiting, and network access control as architectural concerns.

Access Control

Implemented JWT, Redis-backed sessions, IP allow/deny lists, and sliding-window rate limiting.

  • JWT
  • bcrypt
  • Redis
  • forward_auth
Deployment

Deploying and maintaining services with Linux, Docker Compose, Caddy, Nginx, and Cloudflare.

  • Linux
  • Docker
  • Caddy
  • Cloudflare
Secure Networking

Built tooling around WireGuard access, client configuration lifecycles, and gateway authorization.

  • WireGuard
  • VPN
  • DNS
  • Network Boundary
03

Security Analysis & Experiments

Compressing noisy data into explainable results while preserving reproducibility and traceability.

Alert Intelligence

Built alert filtering, sliding-window deduplication, 15-dimensional features, Kafka data flows, and LLM-assisted event review.

  • SOC
  • WAF
  • Kafka
  • LLM
  • Deduplication
Clustering

Built K-Means++ clustering with gonum and selected K automatically using Elbow and Silhouette metrics.

  • K-Means++
  • Elbow
  • Silhouette
  • Z-score
Reproducible Simulation

Studied PoW network behavior with controlled seeds, explicit assumptions, and structured reports.

  • Discrete Simulation
  • PoW
  • DAG
  • Experiment Design
04

Web & Engineering Practice

Moving from rapid prototypes to maintainable delivery with tools chosen for the problem.

TypeScript

Used across Astro, Vue/Nuxt, and Bun/ElysiaJS websites and service prototypes.

  • Astro
  • Vue
  • Nuxt
  • Bun
  • ElysiaJS
Data & Storage

Used PostgreSQL and Redis in services, with Kafka-based streaming security-log integration.

  • PostgreSQL
  • Redis
  • Kafka
Delivery & AI Assistance

Using GitHub Actions, CI/CD, MCP, and AI agents while retaining human verification and engineering judgment.

  • GitHub Actions
  • CI/CD
  • MCP
  • AI Agent

Experience / Field work

Finding operable order in noise and boundaries.

SOC / Security Operations R&D

E01

WAF Alert Intelligence & Triage System

Independent Go Developer

Turned noisy multi-vendor WAF/IDS alerts into traceable, reviewable security events, with Kafka data flow and LLM event review implemented. Reported figures come from one historical offline sample validation and vary with input data, deduplication windows, and configuration.

Delivered

  • Built parallel CSV parsing, five-stage noise filtering, and three-pass sliding-window deduplication in Go.
  • Implemented 15-dimensional feature engineering, K-Means++ clustering, automatic K selection with Elbow/Silhouette, and multi-channel output.
  • Designed a single-buffer LogStore with offset indexes to reduce GC scan pressure while preserving log_id traceability.
  • Integrated offline CSV replay, batched Kafka JSON consumption, and Kafka output, plus optional large-context triage, detailed review, and unified llm_review output.

Outcomes

  • In a historical offline sample, Phase 1 reduced 155,501 alerts to about 2,900 structured events, a 98.1% reduction.
  • Each output event merged 19 source alerts on average and up to 12,000+, while retaining log_id traceability.
  • In a local qwen3.5 review sample, four-way concurrency processed 30 events in about 42 seconds with no errors across that sample.
  • Go
  • Worker Pool
  • LogStore
  • K-Means++
  • Elbow
  • Silhouette
  • Kafka
  • LLM
Technical notes
98.1%Historical sample reduction155,501 source alerts reduced to about 2,900 structured events.
19 alerts/eventAverage consolidationA single event merged up to 12,000+ source alerts.
~42 secondsLLM sample reviewLocal qwen3.5 with four-way concurrency across 30 events.

One historical offline sample

155,501 raw alerts
≈3,000 actionable events
98.1%Historical sample reduction
19 alerts/eventAverage consolidation
~42 secondsLLM sample review

Network Security / Full-stack R&D

E02

WireGuard Secure Intranet Access

Full-stack developer, leading implementation within the assigned scope

Delivered WireGuard-based VPN access for industrial-control hosts on a company intranet, covering client connections, configuration files, and service-state management.

Delivered

  • Designed client connection management and configuration generation, management, and lifecycle flows.
  • Built interfaces for server status and client administration.
  • Validated the prototype with Bun/ElysiaJS, then migrated it to Go/Gin and delivered it with Docker.

Outcomes

  • Produced a public WireGuard UI management tool and repository.
  • WireGuard
  • Go
  • Gin
  • Bun
  • ElysiaJS
  • Docker
Public artifact
Technical notes

Selected work / Built to inspect

Two scales: consensus networks and service boundaries.

Featured work keeps the full technical path and its limitations visible, with live sites and source listed separately.

More work
03

Secure Access Management

WireGuard UI

A management interface and server-side tool for WireGuard access, published as an outcome of the intranet-access system.

04

Multi-agent Web Collaboration

AI Collaborative Web Engineering

A collaborative web project built with peers using the TRAE AI IDE, organized through MCP, editable agents, and structured prompts.

Approach / Working principles

State the assumptions, then make the result reproducible.

  1. 01

    Performance with Security

    Evaluate throughput, latency, memory, and runtime pauses alongside authentication, access control, and boundary configuration.

  2. 02

    Security by Design

    Make security part of the architecture and service boundaries rather than a patch after release.

  3. 03

    Work from First Principles

    Implement or verify critical mechanisms and use small, clear tools to solve practical engineering problems.

  4. 04

    Reproducible & Explainable

    Control random seeds, state parameters, and retain structured output so results can be reproduced, traced, and discussed.

  5. 05

    AI-assisted, Engineer-led

    Use AI to accelerate implementation, refactoring, optimization, and documentation while retaining human verification and accountability.