Technical overview
For IT, data and security reviewers: how DataBridge is installed, what it reads, how it is checked and how you leave.
An enterprise platform that turns your organisation's fragmented data into trusted, usable knowledge that every team can find, check and act on.
The other pages of this site keep to plain business language and point here for the detail.
Architecture and deployment
Two layers on one foundation
A Company Brain that connects your data, and the intelligence layer that puts it to work for every team.
Your environment
- The record store
- Your sources
- The model you choose
- The people who approve
- In the on-premise setup, DataBridge uploads nothing and makes no outbound connection.
- DataBridge reads data where it sits and feeds the tools you choose. Your systems keep running, and you retire an old system on your own schedule.
- The record is updated from the systems you already run, on a schedule your operator sets.
- We set DataBridge up on your machines and connect it to your sources during the pilot.
What stays where
| Item | Where it stays | How you can check it |
|---|---|---|
| Your records and source files | On hardware you control, in your environment | Ask us to run the ordinary work with the network unplugged |
| Connections out, in the on-premise setup | None: it uploads nothing and makes no outbound connection | |
| The model | The model you choose, hosted locally, or none | Matching, reconciliation, reports and exports run with no model |
| Who sees what | By role, set by your administrators | |
| Confidential material | A separate store behind its own key | The everyday view is the same whether or not that store holds anything |
| What a model drafts | It waits for a person, and drafted emails go out only after approval | Shown in the demonstration |
| Your record when you leave | Exported whole in open formats, and restored into a fresh system in front of you | The digest you keep, and the self-test |
Every part of DataBridge in detail
What it is
- Enterprise productivity · Data and AI infrastructure · Runs inside your environment
- We make any organisation's scattered data immediately usable: connected, understood and trusted, whatever its format, source or system.
- Category: enterprise productivity, data and AI. The bottleneck is access to data.
- Lack of AI-ready data puts AI projects at risk.
The five steps
- Connect: reads databases, documents, spreadsheets, email and the systems your teams use.
- Understand: recognises the same customer, supplier, project or person across every source and links them into one record.
- Convert: serves that record in the shape each job needs: answers, reports, analytics tables, AI context and training datasets.
- Prove: each figure read from a file points to the cell, line or page it came from, with its validation status.
- Protect: runs inside your environment with the model you choose.
Inside the Company Brain
- Full history: when each fact was true and when it was recorded, so any past answer can be reproduced.
- Business vocabulary: one agreed meaning for customer, order, contract and project, used by every team.
- Relationship map: how records connect, contract to customer, invoice to order, account to owner.
- Source snapshots: a write-once copy of what each source said when read, so figures stay checkable.
- Change tracking: when a source value changes, every answer built on it is flagged.
The intelligence layer
- Automation and agents: describe the task in one sentence. DataBridge drafts the work and a person approves it.
- Runs inside your environment. Nothing leaves your control.
Security and control
- Your choice of model: use the model you prefer, host it locally, or run with none. Matching, reconciliation, reports and exports need no model at all.
- Runs on your machines: installed in your environment, on hardware you control. In the on-premise setup nothing is uploaded and nothing calls out.
- Access by role: each person sees only what their role allows. Confidential material sits in a separate store, behind its own key.
- A person confirms what AI proposes: facts a model suggests wait for a person to confirm them, and drafted emails go out only after approval.
- Every version is kept: corrections add a new version and keep the old one, so you can see what the record said on any past date.
- Leave with everything: one command exports the whole record in open formats, and we restore it into a fresh system in front of you.
Model options
-
Work that needs no model
Matching records, reconciling, building reports and exporting are ordinary software. The same inputs and rule version give the same result.
-
Reconciliation by rules a person can read
Reconciliation is ordinary software comparing two records of the same money by rules a person can read.
-
The model you choose, or none
You decide which model DataBridge uses, you can host it locally, and DataBridge also runs with no model at all.
-
Nothing a model suggests is added or sent until a person approves
A fact a model suggests waits outside the record until a person confirms it, and a drafted email goes out only after approval. Each confirmation is a dated record.
-
Numbers checked by ordinary code
A number in a model-written sentence is checked by ordinary code against a cited fact, or the sentence is refused.
-
Instructions hidden in files are quoted as content
An email or a file can carry text written to steer a model. DataBridge treats it as content, which can be quoted; in a test on synthetic data in August 2026, 26 of 26 planted attacks were contained, and none of the planted instructions was followed.
-
What is logged for each call
Each use of AI is recorded, with a digest of what went in and what came out.
Data handling and history
Every version is kept
A correction adds a new version and keeps the old one, and an attempt to edit a recorded fact is refused. You can read what the record said on any past date.
Each figure read from a file points to its source
Each figure read from a file points to the cell, line or page it came from, with its validation status. The original files are kept unaltered.
Access control
Access follows each person's role, and confidential material sits in a store of its own with its own key.
Who sees what
Your administrators set the roles once, and access follows each person's role from then on.
Confidential material in a separate store
Confidential material is kept in a separate store with a key of its own, and the everyday view is the same whether or not that store holds anything.
Connectors and formats
Databases, documents, spreadsheets, email and the systems your teams already use. Each source connects once, and your systems carry on as they are.
Your operator decides how each source reaches the record and how often, and every route runs inside your environment.
Twelve kinds of record, and where each is kept today
- Customer records: kept in the CRM.
- Email threads: kept in a shared mailbox.
- Contracts: kept as signed PDFs.
- Ledgers: kept in the accounting system.
- Invoices: kept as PDF and email.
- Spreadsheets: kept on a shared drive.
- Project docs: kept in the team wiki.
- Support tickets: kept in the helpdesk.
- Delivery logs: kept in the operations database.
- Meeting notes: kept in shared folders.
- Handover notes: kept in personal folders.
- Policies: kept on the intranet.
Exports
The whole record can be exported in open formats, and the export is tested in front of you.
Leaving with everything
One command exports the whole record in open formats, with every value, every version and every source reference. We restore it into a fresh system in front of you, the result is checked against its digest, and you keep the digest.
The export self-test
A self-test of the export and restore ships with the product, and your team can run it without writing a line of code.
The shape each job needs
Serves that record in the shape each job needs: answers, reports, analytics tables, AI context and training datasets.
The same value in every shape
An answer, a table row, an export entry and a workbook's value cells carry the same value and the same source as the fact they came from.
What a pilot hands over
You keep the record built from your data and the transcript of everything asked and answered during the pilot. One command exports both, with a digest you keep.
What a security review receives
Your IT, privacy and vendor-risk reviewers get the full architecture sheet before anything is installed. The disclosure policy and the model use statement are on this site.
Security questions and reports go to outreach@walnut.sg.
How this website itself is run
- Every script on this site is served from this domain, and nothing is loaded from anyone else.
- A strict content security policy covers every page, and every connection uses HTTPS.
- The contact form sends your message as one plain-text email to one inbox, and that inbox is where it is kept.
- The contact form asks your browser to solve a small computing puzzle before it accepts a message, which slows automated floods. It uses no CAPTCHA and no third-party service.
- The contact form limits how many messages one network address can send in an hour, and how many the site accepts in total.
- The contact form logs no message, no name, no email address and no network address. It keeps only counts, and a salted hash of your network address that changes every day.
- The site stores up to three choices in your browser: motion, theme and solid surfaces. It stores each one only once you make it.
- Our security contact is published at /.well-known/security.txt.
The model use statement, for your records
- Matching, reconciliation, reports and exports need no model.
- You decide which model DataBridge uses, you can host it locally, and DataBridge also runs with no model at all.
- A fact a model suggests waits outside the record until a person confirms it, and a drafted email goes out only after approval. Each confirmation is a dated record.
- A number in a model-written sentence is checked by ordinary code against a cited fact, or the sentence is refused.
- An email or a file can carry text written to steer a model. DataBridge treats it as content, which can be quoted; in a test on synthetic data in August 2026, 26 of 26 planted attacks were contained, and none of the planted instructions was followed.
- Each use of AI is recorded, with a digest of what went in and what came out.
Why DataBridge wins
Six capabilities, compared across the kinds of tool teams use today.
| Capability | Data warehouses | Search and retrieval (RAG) tools | Feature stores | Large AI platforms | DataBridge |
|---|---|---|---|---|---|
| Connects databases, documents and email | Data warehouses partial | Search and retrieval (RAG) tools partial | Feature stores no | Large AI platforms Yes | DataBridge Yes |
| Links the same customer or case across systems | Data warehouses manual | Search and retrieval (RAG) tools no | Feature stores manual | Large AI platforms services-heavy | DataBridge Yes |
| Serves answers, tables, AI context and datasets | Data warehouses tables only | Search and retrieval (RAG) tools RAG only | Feature stores features only | Large AI platforms custom | DataBridge Yes |
| Every figure traces to its cell, line or page | Data warehouses lineage only | Search and retrieval (RAG) tools chunks | Feature stores features | Large AI platforms platform | DataBridge Yes |
| Runs in your environment with any model | Data warehouses cloud-bound | Search and retrieval (RAG) tools varies | Feature stores Yes | Large AI platforms enterprise | DataBridge Yes |
| Same data for training, testing and live use | Data warehouses no | Search and retrieval (RAG) tools no | Feature stores Yes | Large AI platforms custom | DataBridge Yes |
Only DataBridge spans the Company Brain, the meaning of your data and every output your teams use.
It starts with a demo on a sample of your own documents and data.