About
A small team that ships to production.
MiraiTech is a small engineering team in Nakano, Tokyo. We build AI systems for companies in Japan and abroad, and we only take on work we expect to put into production.
Background
Web and data first, then AI.
The team came together in Tokyo in 2022. Before that we spent years building web applications and data systems for Japanese companies in logistics, manufacturing and finance.
When we moved into AI work, that background mattered more than we expected. The language model is a small part of a working system. The rest is APIs, interfaces, data pipelines, deployment and monitoring, and that is where most projects succeed or fail.
Today we build retrieval systems, agents and LLM features, together with the software they need to run. Every project is led by an engineer who writes code on it, and clients talk to that engineer directly, in Japanese or English.
Principles
Four rules we work by.
- 01
Measure before and after
Every AI feature gets an evaluation set before it gets a prompt. If we cannot measure whether a change helped, we do not ship it.
- 02
Prefer boring technology
Postgres before a new database, a script before a platform. A new tool has to show it solves a problem the old one could not.
- 03
Plan for the model being wrong
Models time out, return the wrong format and are sometimes confidently wrong. We decide up front how each case is handled and where a person checks the result.
- 04
Write things down
Decisions, trade-offs and runbooks are documented as we go, in Japanese or English, so your team can run the system without us.
Timeline
Since 2022.
- 2026
Agents in daily use
Agent systems now run every day inside client teams. We spend about as much time on evaluation and monitoring as on prompts.
- 2025
From single features to agent systems
Most new work became multi-agent systems for support, sales and internal tools. We wrote down the review and handover practice that every project now follows.
- 2024
Opened the AI Lab
We started running experiments between client projects, and signed our first multi-year client: a manufacturer whose visual inspection system is still in production.
- 2023
First LLM feature in production
We shipped our first language model feature to real users. From then on, AI and data work were part of most projects.
- 2022
Started in Nakano, Tokyo
Three engineers with backgrounds in web and data systems set up MiraiTech.
Stack
Tools we use most.
- Models and agents
- Claude
- OpenAI
- Gemini
- open-weight models on Ollama
- LangGraph
- LangChain
- PydanticAI
- PyTorch
- Retrieval and data
- PostgreSQL and pgvector
- Qdrant
- BigQuery
- dbt
- Airflow
- ClickHouse
- Redis
- Product
- TypeScript
- Next.js
- React
- FastAPI
- Node.js
- Infrastructure
- AWS (Bedrock, SageMaker)
- Google Cloud
- Azure
- Docker
- Terraform
- GitHub Actions
- OpenTelemetry