Engineering portfolioJapan

Hi, I'm Akira

AI Platform Engineer. RAG and LLMOps. LLM evaluation and guardrails. CI/CD and observability.

I build AI systems and the delivery paths around them. Retrieval, API integration, automated checks and deployment — from the first useful prototype to something I can operate.

43→65%RAG search success 5,500E2E cases, automated solo −60%deployment time 4+ yrsdevelopment & operations
Explore selected work Where I can help
01 Selected work

Small details. Working systems.

Three open-source projects, with the decisions and edge cases left in the code. A closer look at how I build.

01 / API INTEGRATION

Vertex2OpenAI

TypeScript · Cloudflare Workers · Vertex AI

One familiar API for Vertex AI Gemini.

An OpenAI-compatible gateway for Chat Completions and Responses. The work goes beyond forwarding requests: credential rotation, streaming failures, tool calls and JSON Schema each need an explicit contract.

Engineering focus: compatibility at the boundary, including the failure path.

Read the source — Vertex2OpenAI
02 / RETRIEVAL & DELIVERY

RAG delivery lab

Python · ONNX · Azure AI Search · ACA

From a search query to a controlled release.

Multilingual ONNX embeddings, hybrid and semantic retrieval in Azure AI Search, and structured outputs. The same lab connects the application to container delivery: signed images, SBOMs, canary releases and rollback.

Engineering focus: retrieval quality and deployment discipline in one system.

Read the source — RAG delivery lab
03 / STORAGE & RECOVERY

R2FileBox

Vue · Workers · R2 · D1

File sharing built for interrupted connections.

Resumable uploads in 8 MiB chunks, pickup-session accounting and content deduplication. Retryable cleanup handles the less visible part of storage: keeping metadata and objects in sync after partial failures.

Engineering focus: resumability, state consistency and recoverable cleanup.

Read the source — R2FileBox
02 Engineering expertise

Where I can help

A retrieval problem, an API to integrate, a release process to simplify. I connect application work with the infrastructure and checks it needs.

AI Platform · RAG · LLMOps

Retrieval · Evaluation · Guardrails

PRIMARY FOCUS

I use real traces to improve retrieval, then validate model outputs with schemas and artifact checks. The aim is a useful answer with evidence that can be inspected.

  1. Ingest
  2. Chunk
  3. Retrieve
  4. Generate
  5. Evaluate
  6. Guard
BedrockVertex AIDifyOpenSearchLangFuseMCP

CI/CD & DevSecOps

Pipelines · Supply chain

Reusable release workflows, parallel jobs and caching, with scanning and signed artifacts. I turn repeated manual steps into a delivery path the team can maintain.

GitHub ActionsDockerTerraformHelmTrivyCosign

Cloud & Edge

Multi-cloud · Zero Trust

Public applications at the edge; private tools behind access controls. I work across cloud services, networking and self-hosted infrastructure with clear boundaries between them.

AWSGoogle CloudAzureCloudflareTailscale

Observability & Quality

Telemetry · IaC · E2E

Datadog monitoring as code, LLM traces in LangFuse, and automated checks tied to their evidence. I use failures and runtime signals to decide what to improve next.

DatadogLangFuseSentryPlaywright
03 Systems in operation

From repository to running service

A view of the operating layer. Public endpoints are sampled from your browser; protected services are marked separately and their health is not inferred.

BUILD

vertex2openai-cf

OpenAI-compatible proxy for Vertex AI Gemini on Cloudflare Workers, with streaming, key rotation and tool calling.

TypeScriptWorkersVertex AIOpenAI spec
RUN
SHIELDEDWAF
LobeChatPrivate chat via the proxy
Vercel Edge
Private deployment
BUILD

tokyo-transit-monitor

Rail-delay monitor shipped entirely by CI: scheduled Actions, Pages publishing, chat alerts.

PythonGitHub ActionsPagesWebhooks
RUN
PROBING--
Transit boardStatus page rebuilt on a cron
GitHub Pages
Open
BUILD

r2filebox

Cloudflare-native file sharing with resumable uploads on Workers, R2 and D1.

WorkersR2D1Vue 3TypeScript
RUN
PROBING--
FileBoxHealth endpoint, every sample
Cloudflare Edge
Open
PLATFORM Hybrid self-hosted platform A home lab, NAS and cloud services, with separate access policies for public applications and private tools.
  • Cloudflare Tunnel
  • Zero Trust Access
  • Tailscale mesh
  • DNS · TLS · WAF
AZURE VMStirling-PDF
PROBING--
04 Track Record

Shipped in production

Selected professional work. Client details stay private; the role, scope and observed outcomes stay concrete.

  1. ai-platformOngoing

    AI-assisted E2E test automation platform

    AI Platform / Test Automation EngineerSolo

    • Contract-driven Excel → YAML → Playwright pipeline; AI turns specs into tests behind guardrails.
    • Fail-closed: two environments diffed across 12 checks, hash-linked artifacts, no hand edits.
    ~5,500cases across every screen, automated solo
    PythonPlaywrightJSON SchemaGitHub Actions
  2. rag

    Knowledge-search RAG for a major telecom carrier

    AI / DevOps EngineerTeam of 6

    • Dify workflows routed by document type; OpenSearch index design and model selection.
    • Across 10,000+ PDFs, LangFuse traces showed retrieval design, not the model, was the bottleneck.
    43% → 65%search success, +22pt
    AWS BedrockDifyOpenSearchLangFuseOllamaDocker
  3. ci-cd

    CI/CD and operations automation for a telecom company

    DevOps EngineerTeam of 4

    • Shared GitHub Actions flows reused across teams, with Datadog dashboards and alerts.
    −60%deployment time, via parallel jobs and caching
    GitHub ActionsDockerGKEDatadogShell
05 Credentials

Certifications & languages

AMAZON WEB SERVICES

Solutions Architect – Professional

AWS's top-tier architecture certification.

ORACLE CLOUD INFRASTRUCTURE

AI Foundations Associate

ORACLE CLOUD INFRASTRUCTURE

Foundations Associate

ChineseNative
JapaneseBusiness
EnglishTechnical docs · everyday conversation