AI Sovereignty for Every Japanese Company

Got your own? In-House AI

Stop renting AI. Start owning it.

Anthropic, OpenAI, Google — whose hands is your company's AI really in?
Break free from cloud-AI dependence. Now.

CORE PRODUCT

TOROTAKU Sovereign AI Stack

A 3-layer stack: diso / torotaku IDE / diso-lake

You use ChatGPT. You use Copilot. You use Claude.
It's convenient. Your work moves faster.
You're never going back to the pre-AI world.

— But let us ask you one thing —

Tomorrow's price tag on that AI —
who exactly decides it?

Not your company.
A handful of executives in San Francisco, Seattle and Mountain View.

Chapter 1 — The Crisis

The invisible vassalage
has already begun

"It works fine for us" and "we have sovereignty" are two completely different things.

RISK 01

They set the price. They raise the price.

API fees get quietly revised. There's no "loyal customer discount." The AI budget you set in last year's plan may not survive this year's invoice.

RISK 02

Models get killed off without warning

The model your internal systems were built around last year? Six months from now it gets stamped "deprecated" and disappears. The cost of rewiring your workflows is, of course, on you.

RISK 03

Rate limits stop your business cold

"Service temporarily restricted due to high demand." Behind that polite little message: your quarter-end consolidation grinds to a halt. Your contract review halts. Your business halts.

RISK 04

Your data goes somewhere — you can't see where

Internal docs, customer info, snippets of design work. The screen says "this won't be used for training." But there is no guarantee that setting won't change.

RISK 05

They decide what features you're allowed to use

One day, the AI starts refusing to answer certain topics. One day, the terms of service ban your industry. The foundation of your operations rocks every time someone else changes their policy.

We have a name for this —

The loss
of AI sovereignty

Some people say "but it works fine right now."
Yes. Right now. That's the entire problem.

Chapter 2 — The Structure

This is a structural problem
called "cloud vassalage"

An analogy with electricity.

Using a major AI API is like buying electricity from a utility company. Convenient, efficient, sensible.

Except the AI industry is rapidly approaching a state where there are only three power plants in the entire world.

And those three plants set their own prices, supply terms, and specifications, on their own schedule.

Whatever your company builds with that electricity, in the end it's three companies that hold the switch.

The world's AI "power plants"

Just three companies

Anthropic

San Francisco

OpenAI

San Francisco

Google

Mountain View

Control of pricing, supply and specs

100% on their side

Chapter 3 — The Answer

TOROTAKU's
answer

We are not telling you to roll your own AI from scratch.

We are not telling you to stop using GPT-5 or Claude.
That would be as wasteful as refusing to hire a top-tier outside counsel.

Our proposal

Bring in the strong AI as your "outside counsel."
Run the daily work on "your own AI."

Sovereignty stays with you. The big external models
are summoned only when you actually need them, on your terms.

That's exactly what TOROTAKU's flagship product delivers:
TOROTAKU Sovereign AI Stack.

Chapter 4 — The Product

Three Pillars

Endpoint, agent, observability — sovereignty secured at every layer.

PILLAR 01

diso

AI Gateway

Your company's private "AI power plant"

An in-house AI gateway that bundles multiple open-source AIs (Gemma, GLM, etc.) running on your own servers behind a single endpoint. Employee laptops and apps call your internal diso instead of an outside AI API.

✓Immune to price hikes

✓No surprise deprecations

✓Data never leaves your premises

✓Call it as many times as you want, the bill doesn't grow

Think of it this way — you swap your monthly cloud invoice for an in-house electricity bill.

PILLAR 02

torotaku
IDE

An engineer's right hand

A dev environment that physically tames the AI

Runs as a VSCode extension. Instead of lecturing the local AI, it provides physical discipline and assistance. Our motto: "Code quality isn't decided by how smart the AI is. It's decided by whether it can follow procedure."

FEATURE 01 Physical Guards

Physically blocks "fake thinking," "plan-mode violations," "loops" and "false-completion claims."

FEATURE 02 Hirameki System

Four kinds of "Hirameki" (flash-of-insight) functions that summon a strong AI only at the hard parts.

FEATURE 03 Sub-agents

Investigation feature that runs up to four local AIs in parallel.

FEATURE 04 LSP Integration

Every time the AI reads a file, compile-error info is automatically attached.

BIGGEST DIFFERENTIATOR

FEATURE 05 Intelligent Memory Compression (Condense)

Even on multi-hour continuous tasks, the AI doesn't forget what it just did. We've structurally fixed the "compression accidents" that other tools in this space keep stepping on.

FEATURE 06 Production-grade Hardening

git log freezes, history corruption on resume, 15-minute timeouts — we've stepped on every one of those landmines and patched them.

Push local AI up to a level where it's actually usable, summon strong models as "outside counsel" only when necessary, and we've already detonated every operational landmine for you.

PILLAR 03

diso-lake

Observability & Audit

A platform that keeps the AI's "behavioral record" in your own hands

If you're going to let AI handle real work, you need to be able to record what it did, in-house. diso-lake stores everything on your servers and produces a daily summary report.

•Which model, when, was asked what, and answered how

•Which employee handed which task to the AI

•When the AI got something wrong, why it was wrong, and whether it can be reproduced afterwards

Accountable to your auditors and your internal-controls team.
None of this information is available to you with an external AI API. Period.

Chapter 5 — Proof

Outstanding results on
large-scale mission-critical systems

"Can local AI really do real work?" — we've already answered that.

WORKLOAD 01

Incident Investigation

Tracing the root cause of complex defects, establishing reproduction steps, drafting fix strategies — heavyweight work that normally burns through senior-engineer hours.

WORKLOAD 02

Documentation

Design specs, technical specs, operations manuals, hand-off material — the heaviest knowledge work in mission-critical systems operations.

AI used — both local models

GLM-4.7

LOCAL / Reasoning

Logical reasoning, building cause hypotheses

Gemma-4

LOCAL / Parallel Investigation

Broad investigation, comprehension, summarization, parallel pre-research (multiple instances running concurrently)

With only this combination of two local AIs, we hit production-quality output.
Zero dependence on any external AI API.

¥0

Additional fees to external APIs

Customer-info or internal-doc leaks outside the company

None

Service outages, deprecation notices, price hikes

Yes

Operates even in air-gapped environments

TOROTAKU Sovereign AI Stack is not a hypothetical pitch.

It's a real tool, already producing results on the front lines of large-scale mission-critical projects.

Chapter 6 — The Effect

What happens to companies that adopt this

BEFORE

Your company today

✗Every month, the AI invoice comes in 1.5× what you budgeted
✗You get an email from the vendor: "this model will be retired in 3 months"
✗You're a little nervous about pasting in customer data, so you've quietly limited AI usage
✗One day the AI's output quality changed, you noticed, and you have no idea why
✗The board asks "are we too dependent on AI?" and you can't answer

AFTER

After deploying TOROTAKU

✓AI cost is fixed: server depreciation + electricity + ops headcount
✓No deprecation notices arrive (you own it)
✓Confidential data goes into the AI without leaving the building
✓Every AI action is logged in-house, and quality drift becomes visible
✓At the board meeting you can answer: "Yes, we hold AI sovereignty."

Visualizing the shift

External AI dependenceBefore → After

In-house AI sovereigntyBefore → After

Confidential-data leak riskBefore → After

* Illustrative. Actual ratios vary by deployment scale and operational setup.

Chapter 7 — Timing

Why now

There are three time horizons.

+6 months

Prices will go up. Period.

Cloud AI is in the late stages of "market expansion." Next comes the monetization phase — i.e. price-hike phase. Every SaaS — AWS, Salesforce, all of them — has walked this exact path.

+12 months

The dependence will get deeper

By the time your internal workflows have been rewritten to be unworkable without AI, the cost of switching vendors becomes 10× what it is today. You can only migrate while the dependence is still shallow.

+24 months

Self-hosted AI keeps getting smarter

Open Gemma-class models leap forward in capability every six months. As of 2026, "80% of daily work." By 2027, "90%." Lay the foundation now and every subsequent capability gain lands directly inside your company.

Put the other way around —
companies that don't move now
won't be able to capture next year's capability gains
inside their own walls either.

Chapter 8 — Adoption

A 3-month roadmap
to break free

PHASE 1 / Month 1

Stand up diso in-house

•Leave employee PCs untouched. Bring up diso on a single in-house server.
•Switch the AI API endpoint from an external URL to your internal URL.
•That alone takes data-leakage risk to zero.

PHASE 2 / Month 2

Roll out torotaku IDE

•Distribute torotaku IDE to dev and IT departments.
•Local-AI-powered code assistance goes live.
•Strong-model calls limited to the hard parts only.

PHASE 3 / Month 3

diso-lake audit platform

•Aggregate AI usage logs, generate daily reports.
•Plug into internal-controls and compliance reporting.
•AI sovereignty achieved.

Total: 3 months to a complete roadmap out of external-AI dependence.

Chapter 9 — Why Us

Why companies pick TOROTAKU

Battle-tested on large mission-critical projects

We carry through the heaviest workloads — incident investigation, documentation — using only local AI (GLM-4.7 + Gemma-4).

Entire product built in-house

Not a Frankenstein of foreign open-source projects. We own the design and implementation across all three layers.

On-prem or cloud, your call

Deployable to fit your security policy and infrastructure realities.

We don't blow up your existing AI workflow

Keep using strong models. Just on our turf. The design isn't "throw it out," it's "swap who's in charge of whom."

Designed for Japanese teams and Japanese ops culture

Built by a Japanese dev team that knows what vendor lock-in actually feels like, tuned to the way Japanese companies actually run.

Still evolving, every week

torotaku IDE runs in our own production while we ship improvements weekly. Every issue the front line surfaces, we step on and squash inside the same week. The product is the residue of that loop.

FAQ

Frequently Asked Questions

On a large mission-critical systems project, we have completed incident investigation and documentation work using only the local combination of GLM-4.7 + Gemma-4. The "Hirameki" feature briefly summons a strong model only at the genuine hard spots, structurally compensating for the typical local-AI weakness (raw smarts).

No. We propose "swapping who's in charge," not "throwing it all out." GPT-5 and Claude stay available as your hard-spot-only outside counsel. Call volume drops to under 1/10 of today's, so API bills typically fall significantly.

Today we run primarily on a multi-machine setup with several hundred GB of RAM each. Capacity scales to whatever the user needs.

torotaku IDE runs as a VSCode extension, so existing dev environments are reusable as-is. Anything that speaks the OpenAI-compatible API (Cline, Continue, Aider, etc.) can be redirected through diso by just rewriting the API URL.

It supports fully on-premises operation, so you can use AI without confidential data ever leaving your premises. On top of that, diso-lake stores all AI usage logs in-house and auto-generates daily summary reports — accountable to your internal-controls team and to your external auditors.

No. diso bundles OpenAI-compatible local AI servers (LMStudio, Ollama, etc.), so you can swap in new models as they're released. Capturing every Gemma-class capability bump directly into your own infrastructure is, frankly, one of the biggest reasons to self-host in the first place.

Currently it's a product in active production use inside TOROTAKU and on partner projects. A general external release is planned for the near future. Treat this page as a showcase of what we've built, not as a sales announcement — for now.

— So we'll ask one more time —