HoujunWork with us
Bring AI to your data,
not your data to AI.
Houjun Co., Ltd. is a small AI research and development company based in Japan. We make AI models run privately on your own hardware — Macs, workstations and on-premise servers — and turn them into software people can actually use.
Our lead engineer has worked on intelligent systems since 1998: medical signal recognition, large-scale real-time video, robot vision, and now generative AI running locally.
What we can do for you
Scope, acceptance criteria and a fixed price are agreed in writing before we start. The ranges below are a guide to what similar work usually costs, not a quote.
-
Local AI feasibility review
1–2 weeks · typically USD 3,000–6,000
We test candidate open-weight models on your hardware and your sample data, and measure speed, memory use and quality on your actual task. You get a written report with a clear recommendation — including “don’t do it” if a cloud service or a simpler approach is the better answer.
-
Proof of concept
3–6 weeks · typically USD 10,000–30,000
A working prototype on your data, running where your data lives.
-
Model training and adaptation
Quoted per project
Fine-tuning and domain adaptation of open-weight language models on your own documents. We design the evaluation first, so you can tell whether the model really learned your domain or only became familiar with the test format — and whether it forgot general skills along the way.
-
Licensing and integration
Quoted per project
Commercial licences for VoxStage, and help building VoxStage or Image MLX Lab into your product or workflow.
Our public work
You can download and run all of it before you write to us.
VoxStage
Offline multi-voice audiobook and dubbing workstation for Apple Silicon Macs. Every generated sentence is transcribed back and checked against the script.
houjun.dev/voxstage → Open source · MITImage MLX Lab
Local image generation and masked AI editing on Apple Silicon: only the selected area changes, every pixel outside it stays untouched.
The default model, Qwen-Image, is licensed for non-commercial use; a commercial product needs a licence from Qwen or a different model.
houjun.dev/iml → ResearchIn progressMedAdapt Lab
Training open-weight language models on openly licensed medical papers, on local hardware. We measure what domain training actually changes: new knowledge, tested on held-out papers the model never saw, and forgetting of general language. The training pipeline is running; formal results are not yet published.
houjun.dev/lab →Good fit
- Your data cannot leave your machines: medical, legal, unpublished or customer data.
- Cloud API costs grow with every request, and you want a fixed-cost alternative.
- You need AI that works offline or inside a closed network.
If a cloud service already solves your problem well and cheaply, we will tell you.
How we work
- Written first. We work mainly in writing — English, Chinese or Japanese — so every decision is on record. Calls when needed.
- Your data stays with you. We can work on your infrastructure; NDAs are fine.
- Fixed scope, fixed price, agreed before work starts. Invoices from a Japanese company.
- Time zone: Japan (UTC+9).
Contact
Email support@houjun.dev with the subject “Project enquiry”. Tell us briefly:
- What you want AI to do
- What kind of data is involved
- What hardware you have or plan to use
- Your timeline
We reply within two business days with a written first assessment.