We build ontology platforms.
The part that has to be right before the model matters.

An ontology platform developer. We are a small company in Hadano, Kanagawa, and the whole business is one practice applied three ways: to our own products, to client systems we build hands-on, and to contract development.

The machine-readable form of this table is the fact ledger, /facts.json. It is generated from the same source as this page, so the two cannot say different things.

A retrieval deployment that answers badly is rarely suffering from the wrong model. It is usually suffering from three things that happened years before anyone typed a prompt.

None of that is fixed by a larger context window or a newer model. It is fixed by writing the structure down — which is unglamorous, finite work, and it is the work we do.

Everything below is an instance of that diagram. The products are what happens when we apply it to our own problems; the services are what happens when we apply it to yours.

“Advanced AI work” is a phrase that survives contact with nobody. Here is the concrete version, from Hadano AI Cabinet, the knowledge database we build and publish at h-c.ai.

Many AI memory tools lean on semantic search alone for every query. That is expensive, it is non-deterministic, and hard to audit when it returns something wrong. Hadano AI Cabinet inverts the order: find the document first, then expand only from what was actually found.

Stage one is keyword search over a character-trigram index ranked by BM25. Trigrams rather than words, because Japanese and other CJK text has no spaces — this indexes it correctly without a morphological analyzer in the path. No model runs, nothing goes over the network, and the same query returns the same result every time.

Stage two expands from at most three seed chunks that stage one actually matched. That cap is the entire design. It is what bounds the cost of a query, and it is what makes every result traceable: there is exactly one bridge between the stages, and nothing else crosses it.

The half of stage two that was designed to take vector similarity alongside the graph walk is not connected in this version. It was built against a storage layer that has since been replaced. Asking for vector expansion returns an error rather than an empty answer that looks like a real one. We would rather ship a smaller thing that is honest about its edges.

Underneath all of it is a storage engine we wrote in C, once the first version outgrew the relational layer it started on. Replacing the layer beneath a retrieval system is how quality quietly disappears, and the only way to know it has not is to have kept the old answers before you throw the old code away. We froze a complete run of the old engine first, then compared the new one against it result by result: 0.5800 against 0.5775 on SciFact, 0.2733 against 0.2757 on NFCorpus, with 83.9 and 87.8 percent of the top ten identical. The aggregate held and about one result in six moved, which is the sort of thing we would rather publish than round away.

It also lets us ask whether retrieval still beats pasting the whole corpus into a large context window. What we wrote down before running anything was that retrieval-fed context does not grow as the corpus grows while a pasted corpus does, and that accuracy holds within ten points. It did. Alongside that: 100 facts about companies that do not exist, 100 questions, and a small model answering 88 of the 100 correctly from search results on 1,400 input tokens against 36 of 100 from the pasted corpus on 28,034, with the pasted window right on none that retrieval missed. Those are the friendliest conditions retrieval will ever get, and we wrote the corpus ourselves — the record at h-c.ai sets out the controls and every condition.

MiiHawk runs a mathematical optimization engine over real boarding data to choose which vehicles and how many — fleet composition, not GPS route guidance. Those are different problems, and conflating them is how shuttle-bus software ends up optimizing the wrong thing.

Pasha Kintai reads handwritten Japanese timecards photographed at an angle, on paper that has been in a factory. The hard part was never the OCR call; it was designing an attendance record that stays defensible under Japanese labor law after a machine has filled it in, and an approval trail that a labor inspector can follow.

Specified, designed, built, operated and supported by us. Contract work uses the same stack and the same people, which is the only reason we are willing to describe either one honestly.

An evidence-first knowledge base for AI agents: it finds the passage that backs an answer, returns it verbatim from the stored original, then expands around it. For one agent it runs as an MCP server over standard input and output, with the knowledge in one file on that machine and no network socket open; for a team, an optional process you run yourself serves the same database over HTTP with keys and grants.

v2 is in development and nothing is installable today. We keep built and available in separate columns on purpose, and the record itself is kept at h-c.ai rather than here.

How the retrieval works h-c.ai

Attendance management for small Japanese manufacturers. Photograph the paper timesheet with a phone; AI-OCR reads the handwriting and produces the record. JPY 100 per person per month, tax excluded, and you are billed only for the people who actually clocked in that month — so a quiet month costs nothing.

Covered by Nikkei Construction, Gekkan Soumu Online, VOIX biz and O!Product AI.

Pricing and mechanism pasha-kintai.com

Employee shuttle bus dispatch optimization for factories across Thailand, Myanmar, Vietnam, Indonesia and Malaysia. Boarding data comes from an NFC tap or a QR scan on an ordinary phone, works offline, and syncs later. Reported effect: up to 40 percent less shuttle bus cost, per a customer case published on miihawk.com.

Pricing miihawk.com

The starting price is published: from JPY 30,000, tax excluded, with no setup fee. Every engagement is quoted in writing for an agreed scope before work starts; small builds such as one kintone app have a fixed menu price, and larger kintone and offshore work is quoted by scope and team. Every price we can state in advance is stated below. Japanese IT and DX subsidies frequently apply, and we help with the application.

The third phase is the one that matters and the one most engagements skip. Handover is inside the scope, not after it — the exit condition is written down before the work starts.

Most people who need this kind of work will not find it by browsing. They will ask an assistant, and the assistant will answer from whatever it can fetch. So this site is built to be fetched: no images to interpret, no text trapped inside JavaScript, and every fact available as plain text and as JSON.

Each surface has one job. facts.json holds the current facts and nothing else: names, numbers, dates, prices, product status. llms.txt holds no facts; it says where to read what. Every English page has a Markdown twin at the same address plus index.md, which is the page’s text without navigation or footer. The JSON-LD inside a page describes the entity that page is about, and sitemap.xml lists URLs and their last change. All of them are generated from one internal source together with the pages, so a price that changes here changes everywhere, because there is only one place to change it. Every published claim carries a status, and designed, built and available are three different words.

All of that exists so an assistant can answer on our behalf. Put us to the test: ask Google’s AI Mode what TechJapan LLC builds and what it costs. If the answer comes back wrong, that is our defect rather than yours — tell us what it said and we will fix the source it read.

There is no illustration, icon, stock photograph or logo anywhere in these pages — the only image files on the domain are the favicon in the browser tab and the card social networks show when a link is shared. The diagrams are drawn in monospace characters, which means they are text: an assistant reads them, a screen reader reads them, and they cost nothing to load.

The reading is also deliberately long. A stock photograph of a handshake tells you nothing about whether we can help you, and a page that can be skimmed in nine seconds selects for people who wanted to skim. We would rather lose that visit early than spend an hour discovering the mismatch on a call.

If you have read this far, the filter worked, and we should probably talk.