Pono means true. Conversational intelligence about your organisation's records.
Ask your records a question. Get an answer you can check.
AI assistants that read your documents, spreadsheets and databases, answer in plain language with the evidence behind every answer, and act on your behalf once you approve. It all runs on hardware you control.
- Learns your business by asking
- Every answer shows its source
- Runs on your own hardware
What does our policy say happens when a site misses its monthly report?
A first reminder goes to the site lead after 7 days. If the report is still missing at 14 days, it is escalated to the regional manager.1
1Reporting policy.pdfPage 4, Section 2.3
“…a reminder is issued after seven days. Reports outstanding after fourteen days are escalated to the regional manager.”
Assistants that do the work, not just answer
Your team builds its own assistants and gives each one a job. They work inside limits you set, and nothing irreversible happens without a person saying yes.
Build one in plain language
Start from a template for your field or from blank. Say what it does, choose the documents and tools it may use, and test it as you go. No prompts or models to learn.
They act, with your approval
An assistant can draft and send email through your own provider, read a folder or calendar, and use connectors. Anything sensitive shows a preview and waits for a person.
Several on one question
Mention more than one assistant in a conversation and the work is split between them, each within its own limits.
Every step on the record
In the team setup, what each assistant did, and who approved it, is written to an audit log that shows if anything was altered.
The problem we are solving
Today's reality
What's needed
TodayWhat you know is spread across documents, spreadsheets and databases that no one can ask a question of.
What's neededAsk in plain language, across your documents and your tables, and get an answer.
TodayReports are prepared on a schedule. By the time someone reads one, it can be too late to act.
What's neededAnswers on demand, as a table or a chart, with the records behind them one click away.
TodayCleaning the data takes days every month, and errors are found after the report has gone out.
What's neededData problems are flagged inside the answer they affect, and fixes go to the data manager. The source records are never changed.
TodayThe same indicator is worked out differently by different people, so the figures don't agree.
What's neededEach definition is agreed once, from a cited source, and every figure is computed from it the same way.
TodayNobody is told when something needs attention. Problems are found after the fact.
What's neededSaved rules check new data the moment it arrives and notify the people who need to know.
TodayEveryone sees all of the data, or nobody sees enough of it.
What's neededEach role sees what it should, down to the individual columns that hold personal details.
TodayAI tools guess when they are unsure, and forget what you told them yesterday.
What's neededIt asks when something is not settled, remembers your answer, and gets better the more it is used.
TodayAn AI answer cannot be traced, so experts have to re-check all of it.
What's neededAnswers bring their evidence: the passage from the document, or the records from the table.
TodaySending sensitive records to a cloud AI service is not allowed, and paying for every question adds up.
What's neededIt runs on hardware you control, with models that run there too. There is no charge per question.
It learns your business by talking to you.
Like a new colleague, it asks when it meets something it does not know: what a code in your data means, which date a report should use, what counts as overdue. It shows you the options your data actually holds, with the evidence for each, and you choose.
Your answer is kept and used from then on. If it got something wrong, say so in ordinary words, such as “no, P means pending”, and it changes what it knows and works the answer out again.
So it improves the more your team uses it. A choice it made for you is shown on the answer, never hidden, and anything it learned can be changed by saying so. In the team setup, each change is listed in a review queue with who proposed it, and approved before everyone relies on it.
How many cases did we close last month?
Your cases table has a status column with three codes. Which one means closed?
- C21 of 150 rows
- P37 of 150 rows
- A92 of 150 rows
C. P is pending.
You closed 21 cases in September.
Read “last month” as September 2026 · Change
Remembered: in status, C means closed and P means pending.
The solution in one picture
- 1
Bring your records
Add documents as PDFs and tables as CSV files.
- 2
Check how they went in
See how each document was read. It tests itself with questions from every section, fixes what it can, and tells you plainly what it could not find. What your codes and terms mean is learned, then confirmed by a person.
- 3
Ask anything
In plain language, to an assistant, or to several at once.
- 4
See the evidence
Passages with the supporting quote highlighted, or the records behind a figure. Where numbers are involved, the assistant can calculate them in code and mark which were read and which were worked out.
- 5
Get told what changed
Watchers run your rules on every new load of data and notify the right people.
- 6
Act, with approval
Assistants draft. Any action, such as sending a message, saving a file or updating a record, can be set to wait until a person approves it.
Answers from your documents, checked before you see them
Upload your guidelines, circulars and reports. Pono reads how each one is organised, then tests itself with questions from every section and fixes what it cannot find, before anyone relies on it.
Every answer quotes the passage it comes from, with the supporting words highlighted, so a reader can check it in seconds.
The words your team uses are learned once and used in every table question that follows.
Checked twice, like double data entry
A question that is easy to misread, such as one with more than one condition or one about change over time, is read a second time, independently. If the two readings differ, you see both, each with the number it gives, and you choose.
Answers from your tables also say how your question was read. If that is not what you meant, say so in your own words and it works the answer out again.
Built to ask, not to make things up
Most AI answers sound sure whether or not they are right. For questions about your tables, every question ends in one of three ways: an answer built from your records, a question back to you, or a plain statement that it cannot answer that from your records. For documents, answers are drawn from your passages and point to them.
It asks instead of guessing
When a term, a code or a threshold is not settled, you get a question, not an answer built on an assumption. It never invents a category or a business rule.
The AI never writes the query
For your tables, the model only reads what you asked for. The system builds the query from what your data actually holds, so a figure comes from your records, not from the model.
Answers point to their source
Answers from documents show the passage with the supporting quote highlighted. Answers from tables open to the records behind them.
Numbers are worked out in code
Where an answer involves arithmetic, the assistant can hand the calculation to code and mark which figures were read and which were calculated.
No AI system is right every time. The difference is that this one is built to say when it does not know, and to show you where each answer came from, so a person can check it in seconds.
Runs on NVIDIA, inside your walls.
Pono comes ready to run on NVIDIA GPUs you own. Your records, the search index and the AI models all stay on your machine, so there is no cloud service and no charge per question.
Each organisation has its own installation and its own database. We size the hardware with you once, and install Pono ready to use. Already have NVIDIA GPUs? Pono runs on them.
Start
One NVIDIA desktop AI system or RTX workstation
One office or research team
Team
An RTX PRO workstation for a department
Several teams, shared rooms, watchers
Organisation
A server with NVIDIA data-centre GPUs
Many sites, every engine running at once
No prices here: we quote per organisation after sizing.
Designed for the people who use it
The experts who know the work stay in charge of it. These run in the team setup, which uses sign-in.
Sign-in and roles
Everyone signs in. What they see follows their role, down to individual columns.
Expert corrections, reviewed
When someone corrects how a term or code is read, the change is proposed and approved in a review queue, then used from then on.
An audit log you can verify
Actions are recorded in a chained log that shows if an entry was altered, with a check you can run at any time.
Shared rooms
Several colleagues and the assistant in one conversation, with each member's consent recorded.
Packs for your field
The product is the same everywhere. A pack fits it to one field: the assistants people start from, the vocabulary questions are asked in, and how its data is read and protected.
Healthcare
Health registers and clinical records, starting with maternal and newborn health: clinical vocabulary, analysis of registers loaded as tables, and roles for clinicians, researchers and data protection officers, with personal identifiers hidden by role.
Accounting
Ready-made assistants for tax preparation, reading invoices and receipts, bookkeeping and ledger questions, and drafting client emails.
General
Start from a blank assistant and shape it in plain language, on the same documents, tables and tools.
Another field gets its own pack rather than a separate system. Ask us about yours.
What it works with today
What it does now, and what is still to come.
- Documents
- PDFs with a text layer. Scanned documents, Word and Excel files are not supported yet.
- Tables
- CSV and TSV files, including exports from your databases and spreadsheets.
- Sending through your own provider: Gmail, Microsoft 365, Outlook, iCloud, Yahoo, Zoho, Fastmail or any SMTP server.
- Connectors
- File storage, team chat, accounting software, a folder on the machine, a calendar file, and any tool that offers an MCP server. Each is set up with your own account during installation.
- Models
- Run on NVIDIA GPUs through vLLM, or Ollama for smaller set-ups. We size the hardware with you.
Coming next: live connections to your databases · every question read twice and the two readings compared · data problems flagged inside the answer they affect · photographs of handwritten registers read into data · Word and Excel documents. Each moves up to this list when it ships.
Built on the Aroha protocol
Identities, shared rooms and sharing between organisations use the Aroha protocol, an open protocol (draft specification) that Aroha Labs maintains for keeping AI agents within the limits people set. It is public, so what it does can be checked.
Tell us about your records.
Write to us about the documents and tables you keep, and the questions your team answers from them.