LLMs that keep your data in.
Built for your work.

Post-training know-how from 750B-scale models, for LLMs that run in your closed environment —
and assistive AI built by the people who use it.

Two pillars

Other services: Business AI (RAG, chatbots) / Inclusive LLM Translation

1. Specialized LLMs for closed environments

We train open LLMs to fit your business, your data, and your operating environment — from model selection and data design to training, evaluation, and operation in closed / on-prem environments — so you can use LLMs even for work whose data cannot go to the cloud.

On-premise LLM development & fine-tuning ›

Architecture of a closed-environment LLM: internal data (docs, records, manuals) is used to train, evaluate, and quantize a specialized LLM inside your environment (on-prem, closed network, or cloud), and your users and business apps use it inside the same environment. No data is sent to external AI services.

Process diagram: 1 model choice, 2 data design (synthesize and check), 3 training (SFT, DPO, GRPO), 4 evaluation (task-specific grading), 5 quantization, 6 operation. Weaknesses found in evaluation are fixed by going back to data design.

Post-training & tool use

Specialize open models for your work with SFT, DPO, GRPO, and more. When needed, we go as far as tool use — training and evaluating models that call search, calculation, internal systems, and APIs to get work done (agentic workflows).

Data design & synthetic data

Pipelines that generate training data from your documents and have another model check facts and quality, so training can start even with little data.

Evaluation design

Task-specific evaluation sets and automatic grading, driving an improvement cycle that finds and fixes weaknesses.

Training infrastructure & closed operation

From training under limited compute (FSDP, distributed training, parameter freezing) to quantization, inference optimization, and setups where data never leaves your environment.

Models: open models from 1B up to the 750B scale (Gemma, Qwen, Llama, etc.)
Environment: your own compute (on-prem, closed network, or cloud)
Pricing: quoted individually based on scale and timeline

Tech demo: making a 1B model do exactly what we intend

Chat screen of Iko: a 1B model running entirely on iPhone

Iko is an app that takes a ~1B-parameter model, trains it into a single role through data synthesis, SFT, DPO, long-conversation training, and quantization, and runs it entirely on iPhone and iPad. The same methods scale to larger models.

See the Iko tech demo

Published models: 20+ models on Hugging Face

Talk to us about a specialized LLM

2. Assistive AI built by a disabled developer

Built by a developer who works through a single head-mounted switch. Our CEO, Katsuki Ono, cannot type, speak, or move his limbs; he operates a tablet through one switch mounted beside his head. Through conversation with LLMs, he developed eleven generations of a personalized language model that predicts his own words.

Beyond producing fluent text, we focus on preserving the user's voice and style, transferring intent with minimal input, and being usable in real life, work, and care settings without friction.

  • AAC (Augmentative and Alternative Communication) AI
  • Input-assistance AI for switch / scanning input
  • Sentence prediction and personalized language models — fine-tuning, synthetic data, evaluation, end-to-end
  • On-device AI that runs on smartphones and tablets
  • Disability-led AI design and user research

Model evaluation and dataset design grounded in the lived experience of an actual end-user with severe disability are hard to replicate from outside.

Learn more about Personalized On-Device AI

Track record

  • Large-scale LLM post-training: sub-team leader of a 10-person post-training team (SFT/DPO) in the University of Tokyo Matsuo Lab's medical LLM project, running fine-tuning on models up to 750B parameters
  • LLM development competition: Matsuo Lab LLM Development Competition 2025 — 1st in the preliminaries, 3rd in the finals (team leader)
  • Peer-reviewed research: single-author paper "Reconfiguration of Ability" accepted at ASSETS 2026, the ACM SIGACCESS Conference on Computers and Accessibility
  • Collaboration with Google: fine-tuned Gemini on about 50,000 of his own conversation messages in Google's Project VOICE, which received an iF Design Award Gold Award (documentary / Google Japan blog)
  • Open models: 20+ models published on Hugging Face

Case studies

Research institute: large-scale post-training for a medical LLM

Challenge: build a large language model specialized for medicine.

Role: sub-team leader of a 10-person post-training team, building the SFT and DPO training pipelines.

Approach: combined parameter freezing with fully sharded data parallelism (FSDP) to train models up to 750B parameters under limited compute.

Google Project VOICE: a personalized communication model

Challenge: let people who cannot easily speak or move convey words that are truly their own, with minimal input.

Role: fine-tuned Gemini on about 50,000 of his own messages, built an auxiliary model to synthesize training data, and iterated through eleven versions.

Result: a model whose output his family recognizes as authentically his voice; the project received an iF Design Award Gold Award.

* Client engagements are confidential; we share what we can in individual consultations.

Other services

Business AI (RAG, summarization, chatbots)

RAG, summarization, translation, classification, internal document search, communication-support AI — realistic AI deployments shaped by your objective, budget, and operating environment. OSS models are supported for offline / on-prem deployment.

Plans: from JPY 500,000   Delivery: from 1 month

Sample: an AI receptionist on LINE

LINE bot demo: AI responding in a LINE chat

A bot that runs on LINE. It can also be embedded in your website or other chat platforms.

Try the LINE bot

Inclusive LLM Translation

AI translation combined with review by disabled staff. We deliver short turnaround, high quality, and reasonable pricing in a single workflow.

Formats: PowerPoint and Word
Pricing: Trial JPY 1 / character (first 10,000 characters) / Standard from JPY 4 / character
Lead time: 1 day per 10,000 characters + 3 business days

Public research output and confidentiality

Our team continues to publish research on LLM training efficiency, model analysis, and assistive AI to the extent confidentiality permits. The published record demonstrates capability in model architecture, fine-tuning, evaluation design, and implementation.

Note: Specific client engagements, model details, and ongoing peer-reviewed research are not disclosed due to confidentiality and double-blind review requirements. We are happy to discuss relevant scope and capabilities under NDA.

Contact

For AI development, LLM R&D, or assistive AI inquiries, please use the form below. Briefly include your project area, timeline, and deployment environment if possible.