FLARESTONE

AI-driven development adoption (FDE) / Generative AI business PoC

Less demo. More delivery.

AI agents handle implementation and review, and tests verify the quality. We bring this development setup, which we run every day in our own company, to your team.

Book a free 30-minute consultationSee how we build

No Microsoft account needed: book as a guestNo sales calls

4,200+Pull requests merged in 2026
50+ parallelTest environments we run ourselves
2productsBuilt with this setup, deployed or being deployed
18yearsOur founder's software development experience (mainly embedded and automotive)

Figures are measured in our own product development (as of October 21, 2026). They do not mean you will see the same figures in your environment. The graphic is illustrative.

PROCESS

People decide, AI builds, the system verifies

In your team, what to build, the acceptance criteria, and the final decision on merging changes are decided by your staff and our founder. AI agents handle implementation and review, and tests check the work against the criteria that were set.

How we split the work when we bring this into your team

  1. 1

    Specification and acceptance criteria Your team + our founder

    We decide first what to build and what counts as done.

  2. 2

    Implementation AI

    AI agents implement each task and propose a change.

  3. 3

    Review AI

    Another AI, separate from the one that implemented, checks the work against our review rules. Each issue to fix is logged one by one, and nothing is merged until all are resolved, or while anything remains unconfirmed.

  4. 4

    Testing and pre-merge checks System

    The relevant tests run in parallel, and nothing is merged if even one thing newly breaks.

  5. 5

    Final decision and exceptions Your team + our founder

    People make the final decision on whether to merge. When the spec is too ambiguous for AI to settle, or a review keeps failing, people revisit the acceptance criteria.

SERVICE

Two ways to start, depending on where you are

If you have a development team, FDE brings in the whole setup. If you first want to confirm whether generative AI works for your business, a 2–4 week PoC.

For companies with an in-house development team

AI-driven development FDEWe join your development team

Our founder joins your development team from two days a week and sets up, together with you, a development process run by AI agents.

Designing and introducing the setupWe build the specification → implementation → review → test → integration flow into your repositories and CI.
Delivering the first resultsWe pick one real task or feature and build it all the way through with this setup.
Managing permissions, records, and costWe define what AI may touch, how records are kept, and the cap on AI usage fees, and operate accordingly.
Handing over the proceduresWe leave runbooks and decision criteria so that your own members can run it on their own.
From JPY 500,000 / monthFrom 2 days a week, excluding tax3 months recommended, 1 month minimum, then month by month
For companies that want to start small (no development team needed)

Generative AI business PoC2–4 weeks

We pick one business task and confirm whether generative AI really works for it. We hand over something that runs, along with the material for deciding whether to proceed to full development.

Narrowing it to one taskWe first set the acceptance criteria: what goes in and what must come out for it to count as a success.
Running it on real dataUsing the data you entrust to us (or an equivalent sample), we build something that works, down to the screens and processing.
Compiling the material for a decisionWe hand over the accuracy, cost, and operational issues, along with an estimate for proceeding to full development.
From JPY 500,000Fixed fee, excluding tax; quoted according to the task and data

If the PoC results are good

You can move on to the next stage

The code built in the PoC is handed over to you.

Full development (contract development) and operations support　We build it and support its operation after launch
FDE　We bring this whole development setup into your development team
STEP 1Free 30-minute consultationOnline; we hear about your current situation and goals
STEP 2ProposalWithin 3 business days, we send the scope, period, and cost
STEP 3StartWe can start as early as 1 week later

CODE & DATA

How we handle code and data

Where things are kept, where they are processed, and who can touch them are agreed in writing before we start.

Nothing taken out to usFDE

FDE work is done on your repositories and CI, and we do not take code out to our own environment.

What external AI processesAI agents

When AI agents (such as Claude Code and Codex) are used, the code and instructions needed for the work are processed on each AI provider's servers. The tools, contracts, and settings (such as not using inputs for training) are listed and agreed before we start.

What cannot leave your companyCase by case

Code or data that cannot be sent outside your company can be left out of scope, or we can discuss a setup using a model that runs on your own servers.

Data entrusted for a PoCPoC

If it is handled in our environment, the scope of the data, where it is stored, and the deadline for deleting it (within 30 days after the end) are agreed in writing before we start.

Permissions and recordsAll services

AI is given only the permissions it needs, and a record of the work is kept.

WORK

What we built with this setup

All of these are built with the development setup above.

Deployed on the client's in-house servers

Business risk analysis system for a manufacturer

Problem
Collect and organize business risk information from public sources. The data used for analysis cannot leave the company.
Scope
Collection from about 47 public sources, risk classification in 9 areas, LLM analysis, and a dashboard. The LLM runs on the client's in-house GPU server and uses no internet connection for inference.
Period
Development started in June 2026, delivered in September, deployed on the client's servers in October
Team
Our founder and this development setup (implementation, review, and testing by AI agents)

The client's name and screens are not shown for confidentiality.

Delivered

PoC of a driver monitoring system (DMS) using an in-vehicle camera

Problem
Detect when the driver looks away or gets drowsy, as well as facial expressions and pointing gestures, from in-cabin camera video.
Scope
State detection from face and hand landmarks (with state transitions that suppress false detections), camera video capture, recording and timeline playback, calibration for individual differences, cloud deployment, and API specifications in Japanese.
Period
Development started in June 2026, delivered
Team
109 pull requests merged with this development setup

Try the demo in your browser →

The names of the client and the co-development partner are not disclosed for confidentiality.

Our own servicePreparing deployment

AI processing of business forms and invoices: ClarityFin

Problem
Invoices and other business forms are checked by manual entry and by eye.
Scope
Reading, digitizing, and searching business forms, and checking the required items of qualified invoices.
Period
Development started in April 2025
Now
We are deploying it for our first paying client.
In-house developmentPrototype

In-vehicle AI agent MIKOTO

Problem
Build an agent that can safely handle everything from in-car conversation to vehicle operations such as the air conditioner.
Scope
Reads the intent of what is said and converts it into calls to 24 kinds of vehicle operations (Function Calling). Only calls that pass a safety gate checking the vehicle state, such as speed, gear, and parking brake, are executed, and the design stops when in doubt. Generates the reply and performs the vehicle operation in parallel, and responds in Japanese. APIs for the cloud (REST and SSE). Intent reading currently uses an LLM in the cloud, and we plan to switch to a small language model on the vehicle side.
Period
July 2026 (13 pull requests merged in 2 days with this development setup)
Background
Developed as our own product, based on experience from contract work from September 2024 to June 2026 (concept and prototype of an in-vehicle agent, and fine-tuning of a small language model).

The name of the contract client is not disclosed for confidentiality.

PROFILE

Masaki AkiyamaFounder & CEO / Head of Development

Our founder handles everything personally, from the first consultation to development. Masaki has 18 years in software development, mainly embedded and automotive, with hands-on work in machine learning (from 2016) and generative AI (from 2020).

M.S. in Electrical EngineeringPMP2 patentsSIGNATE MasterKaggle Bronze Medal
  1. 2008Embedded software developmentUSB drivers, RTOS, and Linux kernel porting
  2. 2010Joined the Japanese subsidiary of a major software companyTeam leader for Linux driver development for customer-specific devices
  3. 2016Advanced technology development at the same companyMachine learning (path planning, pedestrian risk detection, and more)
  4. 2020Department manager at the same companyLed mass-production development of automotive infotainment (3 domestic teams plus an offshore team of over 60). Patented an automated testing tool
  5. 2023Automotive parts manufacturerManager of mass-production projects
  6. 2024Founded FLARESTONE Inc.Development of automotive AI agents and small language models (contract work), and of our own products

TERMS

Contract and fees

All amounts exclude tax. We quote according to scope and number of days.

FDE
From JPY 500,000 / monthFrom 2 days a week
PoC
Fixed fee from JPY 500,0002–4 weeks
Full development
QuotedContract development and operations support
Term
3 months recommended, 1 month minimumAfter that, continue, scale down, or end month by month
Working style
Mainly fully remoteWe visit when needed, such as at kickoff (Tokyo area)
Contract
A service contract with FLARESTONE Inc.We can sign an NDA
Rights
Deliverables belong to youGeneral-purpose tools we already had are licensed for your use
Available from
November 2026
Lead time
As early as 1 week30-minute consultation → proposal within 3 business days

FAQ

Frequently asked questions

Is the quality of AI-written code acceptable?

We set the acceptance criteria first, and merge only what passes review by a different AI and the tests. Anything the criteria cannot decide is decided by a person.

Our internal rules restrict the use of AI.

We first confirm which tools and scope are allowed, and build the setup within them. Anything that cannot be sent outside your company can be left out of scope, or we can discuss a setup using a model that runs on your own servers.

Can we hire you without a development team?

Yes. We start with a generative AI business PoC and, if the results are good, take on full development (contract development) and operations support.

What is the minimum term, and what if we want to stop midway?

We recommend three months for FDE, but you can contract for as little as one month. After that, continuing, scaling down, or ending is decided month by month. A PoC takes 2–4 weeks.

Who owns the code and deliverables?

They belong to your company. General-purpose tools we already had (templates and scripts) are made freely usable by your company.

Can you work on site?

We work mainly fully remote, but can visit when needed, such as at kickoff (Tokyo area).

Start with 30 minutes: tell us where your development team is today.

There is no need to send confidential materials. The size of your team, the languages and tools you use, and what you want to try with AI are enough.

Book a free 30-minute consultationEmail us: [email protected]

No Microsoft account needed: book as a guestThe booking page shows the name of the person in charge (our founder)No sales calls