We build software and the teams that ship it.
An engineering studio in Minato, Tokyo. We build full-stack products and the automation around them, and we run expert data operations on the platforms that train frontier models.
- Years senior delivery
- 10+Years senior delivery
- AI training platforms
- 8AI training platforms
- Industries served
- 6Industries served
Three disciplines, one system.
Most engagements are a mix of all three. They are separated here because they fail in different ways, and knowing which one is failing is most of the work.
Architecture
Decide the expensive things on purpose.
Schema shape, service boundaries, what is synchronous, where state lives. These are the decisions that are cheap in week one and ruinous in year two, so they get made deliberately and written down with the reasoning intact.
Every project ships with an architecture note listing the decisions we would not want you to reverse without reading it first.
Analytics
Cut the system open and look inside.
A product you cannot measure is a product you cannot improve. Events, traces, and one agreed definition per metric — so that when two dashboards disagree, there is a fact to settle it rather than an argument to have.
Instrumentation goes in with the feature, not in a later phase that never gets funded.
Automation
Stop paying people to be a script.
Deploys, checks, migrations, reports, and the manual work sitting in someone's inbox. If a task is repeatable and being repeated by a person, it is a candidate — and if automating it would cost more than the task, we will tell you.
We report automation as hours returned per week, measured against the manual baseline.
Five kinds of work.
Each one is a full engagement rather than a bolt-on. The legend key is how a piece of work is referenced in a scope document.
- FS
Full-stack development
Web applications, APIs, and the infrastructure they run on.
- TypeScript
- Next.js
- Node
- Postgres
- AI
AI automation
LLM agents and retrieval systems wired into work you already do.
- Claude
- OpenAI
- pgvector
- LangGraph
- CMS
Content platforms
Headless CMS builds that marketing can run without filing tickets.
- Sanity
- Payload
- Next.js
- Cloudflare
- CRM
CRM systems
Pipelines, automation, and reporting your sales team will actually use.
- HubSpot
- Salesforce
- Postgres
- dbt
- AITP
AI training platforms
Expert data operations on the platforms that train frontier models.
- Outlier
- Snorkel AI
- Toloka
- Alignerr
Industries served
6 sectors- Fintech
- Ledgers, reconciliation, KYC flows, and audit trails that hold up under review.
- SaaS
- Multi-tenant products, usage metering, billing, and self-serve onboarding.
- E-commerce
- Catalogue, checkout, fulfilment integrations, and traffic spikes that do not take the site down.
- Healthcare
- Patient-facing tools and internal systems built around consent and access control.
- Manufacturing
- Shop-floor data capture, scheduling, and reporting on top of systems from 1998.
- EdTech
- Course delivery, assessment, cohort analytics, and accessibility that passes audit.
How an engagement runs.
Three stages, in order. The numbering is not decoration — each stage depends on the one before it, and skipping the first is how projects fail expensively.
- 01
Listen first
Two weeks before anyone writes production code.
We read the existing system, sit with the people who use it, and write down what we found — including the parts of your brief we think are wrong. You get a scope with named risks and a fixed price for the first milestone. If the honest conclusion is that you should not build this, you get that instead, and you keep the document.
Output · Scope, risk list, milestone price
- 02
Build in the open
Weekly demos against a board you can see into.
You are in the repository and the tracker from day one. Every week there is something running to look at, not a status update. When an estimate moves, you hear it that week with the reason attached — never in a summary at the end of the month.
Output · Weekly demo, open board, running system
- 03
Leave it better
The exit is designed at the start.
Handover is a deliverable, not a favour. Documentation aimed at your next hire, a recorded walkthrough, and a period where your engineers drive while we review. Success is your team shipping the second version without us.
Output · Docs, recorded handover, supported transition
What is on the board right now.
Three active projects is our ceiling. If the next slot is two months out, you will hear that rather than a start date we cannot keep.
- WK-11Fintech
Ledger and reconciliation
Double-entry ledger with nightly reconciliation against two payment providers. Replaced a batch job that had grown from twenty minutes to nine hours.
Nightly run: 9 h → 4 minIn build - WK-12Manufacturing
Shop-floor scheduling
Scheduling and data capture across three plants, reading from a control system that predates the company's current network. Tablets on the floor, one source of truth behind them.
Manual re-entry removed at 3 sitesIn build - WK-13SaaS
Internal automation platform
Agent workflows over support tickets and contracts, with approval gates on anything customer-facing and an evaluation suite running in CI.
≈ 22 h/week returned to the teamIn build
What we have written down.
Method notes and field reports from work in progress. No announcements, no hiring news.
- Method
Write the evaluation before you write the agent
Every AI project we have rescued had the same missing piece: no way to tell whether a change made things better.
- Field report
A reconciliation job that took nine hours, and why
A fintech client's nightly reconciliation had grown from twenty minutes to nine hours. The fix was not more hardware.
- Note
An estimate is a forecast, so give it a range
Single-number estimates are a promise nobody can keep. Ranges are honest and, oddly, easier to sell.
Tell us what you are trying to ship.
Send a short description of the problem. You will get a reply from an engineer who read it, not a form response — usually within one business day.