AI Architect & Technical Advisor

ORNEW's image character: an illustration of a person with teal hair

Not bare AI,but AI as a product.

A model on its own does not move a business. I finish the data, evaluation, inference platform, integration into the business and operations as one product.

Measured bybusiness results,not accuracy.

What makes AI good is the business outcome, more than model performance. I decide what to build by working back from the return on investment.

Built to keep running,built to be stopped.

In production used by tens of millions: it stays up, does not break, and can be rolled back. I design AI that records the grounds for its decisions, on the premise that people can stop it and roll it back at any time.

How I workGet in touch

How I work

Not bare AI,but AI as a product.

What is missing

The gap between research and production is nobody's job.

Researchers watch model accuracy, engineers watch system uptime, and the business watches its KPIs. The evaluation, operations and accountability that keep AI running in production belong to none of them, and proofs of concept never reach production.

What is needed

Not AI itself, but a business system that keeps running.

The business numbers move. It does not stop, and it can be stopped at any time. Every decision can be explained. The organization can keep it when people change. Adding AI is not the goal in itself.

Solution

Business value, created with AI as a product.

How bare AI and AI as a product differ
Bare AIAI as a product
ValueAccuracy went upThe business numbers moved
EvaluationIt worked in the demoMeasured on production data, continuously
StructureOne LLM does everythingLLMs, statistics, machine learning and algorithms, each in its role
UptimeRebuilt after it goes downRedundancy and graceful degradation keep the business running through failures
ControlAI output flows straight into the businessPeople approve it, and can stop and roll it back any time
ContinuityNobody can fix it once its builder leavesThe organization can keep it

Not the model alone: I design the data, evaluation, inference platform, integration into the business and operations as one system, against three standards. It is measured in business numbers, it is composed of roles rather than resting on a single LLM, and it does not stop yet can be stopped.

Role

An engineer who leads the business.

I design AI from the viewpoint of the business, not as a feature. Before deciding what to build, I decide where in the business it creates value. The technical decisions and the business decisions are made by the same person.

Engagement

From a technical advisor a few hours a month to a one-off consultation.

As a technical advisor, I join design decisions and reviews from a few hours a month. One-off consultations are welcome too. All work is fully remote.

  1. Consulting
  2. Research
  3. Development
  4. Integration
  5. Operations

I can join at any stage and design everything from there on.

See engagement terms

Work

Fully automated manga localization

CyberAgent AI Business Unit　2024–

Most manga can only be read in Japanese because of cost, and pirated editions have filled the gap. Machine translation did not understand the work, and typesetting damaged the intent of the expression, so neither reached the quality of an official edition.[1]

I built a fully automated localization system that integrates an understanding AI, a cultural translation AI and a typesetting AI, understands a work deeply and approaches fan-made quality. It runs as the business foundation for multilingual distribution that returns revenue to creators and protects their rights. I am responsible for the three AIs from development to operation, as the person accountable for business value.[1]

Manga manuscriptUnderstanding AICultural translation AITypesetting AIMultilingual editionManga manuscriptUnderstanding AICultural translation AITypesetting AIMultilingual edition

Sources

  1. [1]Understanding manga with AI and fully automated localizationSpeaker Deck, 2025-04-24

Managing data quality as code

LY Corporation AI Platform Dept.　2022–2024

Machine learning knows what the data means, and operations runs the platform. Data quality problems happened in the gap between the two.

I developed a quality definition language, DQML, so that ML and Ops can manage quality as code in the same words. By inspecting every data source automatically, ML engineers can focus on model development and the operations side can guarantee data quality. The platform is called ACP Data Quality.[2][3]

Sources

  1. [2]LY Corporation Tech Blog2023-12-25
  2. [3]Data-Centric AI Study Group, 7th meetingSpeaker Deck, 2024-03-29

Reading LLM safety from internal representations

CyberAgent AI Business Unit　2025

How does an LLM judge the safety of anime and manga expression? I extracted internal features with Gemma Scope 2, visualized them and verified the result.[4]

Sources

  1. [4]Visualizing the internal representations of an LLMCyberAgent AI Business Unit note, 2025-12-26

Profile

Arata Furukawa古川 新

I am an engineer who designs the systems that keep AI running in production. After DMM.com and LY Corporation, I now lead new AI businesses at CyberAgent. An idealist and a realist, I never change one thing: changing. I make this site, the logo and the icon myself.

Read more about Arata Furukawa

ORNEW's image character: an illustration of a person with teal hair

Engage

As a technical advisor,from a few hours a month.

The core of what I take on is technical advisory. I carry the technical roadmap from planning to delivery, the technical management of development teams, and design responsibility for MLOps platforms, on an ongoing basis. Design decisions and reviews start from a few hours a month, all fully remote.

Spot consulting and reviewI also take one-off consultations and reviews: a second opinion on AI adoption strategy, architecture design reviews, and conducting hiring interviews and assessing technical skill.

Validation and launchI sometimes help validate new technology or launch demanding systems: a Go and gRPC backend, a Kubernetes Custom Controller, prototyping an LLM Agent, and the like.

Talks and writingI also take speaking at conferences, writing and contributing technical articles, and leading internal study sessions and lectures.

Please send your inquiry by email.

Discuss by email
Status
Open
Load
From a few hours a month
Form
Fully remote

Not acceptedWork or companies that compete with my current employer. On-site or full-time work. SES or person-month staffing.

Career

  • 2024–CyberAgent, Inc.Business CTO and product manager at the AI Business Unit, leading new generative AI businesses. Responsible for manga localization from development to operation.
  • 2022–2024LY Corporation (formerly Yahoo Japan Corporation)Built the company-wide AI platform at the AI Platform Dept. Led the data quality platform ACP Data Quality as its product manager.
  • 2017–2022DMM.com, LLCOwned AI/ML models from development to production. Built an MLOps platform and improved recommendation and more.

Personal open source

zeroparse

zeroparse/zp, Go. One grammar, every parser

  • protoc-gen-cueA protoc plugin that generates CUE from Protocol Buffers
  • protoc-gen-pydanticA protoc plugin that generates Pydantic BaseModels from Protocol Buffers
  • hffHuman-Friendly Filter Language
  • zerositeA minimal site generator
  • envoy-xfccA Go library that parses Envoy's X-Forwarded-Client-Cert header

Notes

Read all articles

Talks and writing

Talks before 2023

Contact

old.river.new@gmail.com

I may not be able to reply to every email.