Skip to content

We build apps, websites,and the AI inside them.

A product studio that ships client software, and runs nine AI tools of its own, including a knowledge base that runs with a local model and cites its sources.

Start a projectSee our products
  1. 01Your material
  2. 02Index
  3. 03Model
  4. 04Output
  5. 05Check

The model proposes. The code says whether it counts.

RetrieveGenerateVerifyLocal or cloud

What we build on

How we build it.

Most engagements use one or two of these. The rest of what we sell is on the products page.

AI systems

Retrieval, agents, evaluation, and the unglamorous plumbing that keeps a model honest in production.

Design and front end

Design systems and interfaces drawn by the people who also ship them.

The system this page is built on
  • Palette lockOne neutral, one accent
  • Shape lock16 / 10 / pill
  • Rhythm lock72px between sections

Hover a lock

Infrastructure

Deploys, observability, and model spend you can read on one page.

Nine tools we ship under our own name.

Built on the same stack we hand to clients, and available on the web and in your pocket. One at a time.

All products
HangRui IndexHangRui PagesHangRui ScreensHangRui LyricsHangRui VowsHangRui ScoreHangRui StillsHangRui SequenceHangRui Desk

The assistant on this site, one real run

Recorded September 2, 2026; the counts are from that run

Can HangRui Index run with a local model? Does our data leave the building?
Ask

Can HangRui Index run with a local model? Does our data leave the building?

Yes. HangRui Index can run with a local model on your own hardware, so the documents never leave your environment 1. If you use an outside API instead, only the few selected pieces are sent — never the whole document 1.

How it answered

  1. Retrieve5/37The site is cut into pieces and every one is scored against your question. Only the highest scoring few are put in the prompt; the model never sees the rest.
  2. Answer1/5The model may answer from those pieces and nothing else. What it knows from anywhere else does not count here.
  3. Verify1/1Every citation it wrote is matched back against what was actually sent. A number it invented is left as plain text instead of becoming a link.
  • Answered from
  • [1]HangRui Index100%
  • [2]Who else touches your data68%
  • [3]HangRui Pages65%
  • [4]Nine tools, one way of working.64%
  • [5]Do you do custom or on premise work?60%

Every number in the answer links to the page that sentence came from.

Ask one yourself

Knowledge

HangRui Index

Your documents, your machine, answers with sources.

Cut your company's documents into an index, connect a model, and get answers that say which page they came from. The model can be one running on your own hardware, so the documents never leave your environment.

Read more

1 of 9

Retrieval-augmented generation

Answers with sources. Data that never leaves.

We build AI assistants one way: the knowledge lives in an index you keep, and the model only reads the few pieces that were chosen. So the model can be swapped, and it can be the one racked in your own server room. The assistant on this site takes the same path.

The path one answer takes

  1. 01Your documents

    Cut along their own structure, each piece remembering the page it came from.

  2. 02Index

    Every piece is scored against the question. Only the best few go on; the model never sees the rest.

  3. 03ModelOn-premiseCloud

    An open-weight model on your own GPU, or an outside API. Changing supplier is one line of configuration.

  4. 04Answer

    Every citation is checked back against what was actually sent. An invented one stays plain text.

Only the model stage knows who the supplier is; the other three are code. The assistant on this site takes exactly this path on every answer.
  1. 01

    Knowledge in the index, not in the weights

    Change a document, rebuild the index, and the next answer is current. No training run and no regression suite, because nothing was ever learned into the weights.

  2. 02

    The model is a slot

    Of the four stages, only this one knows who the supplier is. Plug in an open-weight model on your own GPU and the documents never leave your environment.

  3. 03

    Citations are checked, not asserted

    Every number in an answer is matched back against the pieces that were actually sent. One that matches nothing stays plain text. What you see was counted.

Meet HangRui IndexAsk this site first
Read why we build it this way

Ask this site.

This is how we build a retrieval grounded assistant: the site is cut into pieces, scored against your question, and only the best few reach the model — then every citation it writes is checked back against what was actually sent. The counts below are from the run you just triggered, not an illustration of one.

Or try

Answers come from a model reading this site’s own pages. It is not a person, and it knows nothing that is not published here.

How a build actually runs.

Four moves. You can stop after any one of them and still own everything we made.

Scope

One week. We map the product, cut what does not earn its place, and price the build before you commit to it.

Prototype

A working slice in front of real users while the full build is still cheap to change.

Build

Two week increments, deployed continuously, with your team in the repository the whole time.

Operate

We stay on after launch for monitoring, model spend, and the next release.

A few things we have shipped.

Client names are under agreement; what was built is not. This section lists only work that was actually delivered, which is why it is short.

Web platformDesign

Web design and front end

Most of the work has been this: a site taken from layout and content model through to the front end, sized for a desktop and a phone, and handed over so the client can edit it themselves.

LINERetrieval

A LINE bot that finds the record

Customers ask on LINE, it finds the matching row in the database and tells them where it is. There is no model in it: the matching and the ranking are code, so it either finds the record or says plainly that it did not. Desk and Index grew out of it.

How we think about this.

MethodMeasurementReproducibility

Reviews of published work with their sources attached, and notes from building our own. Every number here is either measured or cited.

All writing

Two ways to work with us.

Both start with the same one week scope, and the scope is the same price either way.

Project build

A defined scope, a fixed timeline, and a price agreed before anyone writes code.

Best for a first version, or a rebuild that has been put off too long.

  • Fixed scope and price
  • Named team for the whole build
  • Handover with the repository

Embedded team

Two to four of us working inside your team, month to month, in your tools.

Best when the roadmap already exists and there are not enough hands.

  • Monthly, cancel with notice
  • Your process, your standups
  • Scales up and down by the month

Tell us what you are building.

Send a paragraph. If it is a fit you will get a reply from the person who would run the build, usually within two working days.

hello@hangruiai.com

A paragraph is plenty. What it is, who it is for, and when you need it.