In-House Insights

AI analytics that keeps data in-house.

In-House Insights is a sovereign AI analytics platform. It runs on a server the organization owns, answers questions from the databases, spreadsheets and documents the organization already holds, and retains every proven analysis as an asset the organization keeps. The AI model is a replaceable component. The data does not move.

How we engage

Video · two minutes

An introduction from the foundersVideo coming soon

A two-minute introduction: what the platform is, how data remains on your infrastructure, and what an organization retains from its use.

Deployment

A server the organization owns. Users connect from a browser over the corporate network; no client software is installed.

Data

Remains on the organization’s infrastructure. The model receives a description of the data; the analysis runs on the server.

AI model

Claude, GPT, Gemini or Grok on the organization's own keys, or a local model for fully offline operation.

Retained

A reviewed Knowledge Base; dashboards, KPIs, Live Monitor and scheduled reports that run without model calls.

Governance

One administrator; user groups with access denied by default; a 90-day activity record.

Who it is for

Organizations whose most valuable data may not go to an outside service.

An analytics team at work inside a regulated organization: two analysts at a workstation reviewing a spreadsheet and chart while a colleague looks on

Banks, hospitals, ministries, law firms, manufacturers and research institutions have analytics teams, established data and dashboards. They also have a policy, a regulator or a residency law that requires their most sensitive data to remain within the organization. In-House Insights is designed for that situation.

What they have

Databases, spreadsheets and documents; analysts who know the data; dashboards for the figures reviewed every week. Questions outside the dashboards are answered by analysts on request.

What the platform adds

AI-assisted analysis of that data within existing policy, on infrastructure the organization controls. Analyses that have been reviewed are retained and can be re-run without further model calls.

The platform

Three properties of the architecture, and the capabilities built on them.

The model works from a description

Datasets, files and documents are not uploaded to the model. It receives a description of the data, including its structure, summary statistics and a small number of example values at default settings, and writes the analysis; the analysis runs on the organization's own server against the actual figures. This applies to each supported cloud model and to a local model alike.

Reviewed analyses are retained

An analysis that has been reviewed is retained as executable work the organization owns. It can be re-run on current data, used in dashboards, KPIs, Live Monitor and scheduled reports, and shared with colleagues who do not write code, without further model calls.

One server for the organization

Installed once on a server the organization owns. Users work in their own workspaces from a browser; departments are managed as user groups; one administrator governs the installation. No cloud tenant is required.

Over time

As the Knowledge Base grows, a larger share of routine analytics runs from retained work rather than from new model calls, and the organization's dependence on any single AI vendor decreases.

AnalysisPlain-language questions

A question in plain language returns charts, tables and a written explanation from the organization's databases and files. Follow-up questions continue the thread, and the code behind each result is available for review.

DashboardsComposed from a description

A dashboard is composed from a question and supports cross-filtering, slicers, time windows and drill-down. Refreshes run without model calls; boards export as branded reports and can be emailed on a schedule through the organization's own mail server.

KPIsMeasured on a schedule

A KPI is defined in a sentence, measured on a schedule and kept with its history. Alerts are raised when a value crosses a limit or departs from its usual pattern for the day of the week. Reports and scheduled email are included.

Live MonitorNo model calls at runtime

Saved analyses refresh at intervals from one minute to one day and stream to an operations screen and to authorized colleagues.

Knowledge BaseReviewed and permissioned

Reviewed analyses are promoted into collections by department. Authorized colleagues run them on current data without direct access to the underlying tables.

DocumentsIndexed on the server

PDFs, including scanned documents, are indexed on the server and answered with page references, in English, French or Turkish. Findings can be carried into data analysis.

SpreadsheetsImported once, repeated

A workbook with title rows and merged headers is imported after one review; the same layout imports identically in later months.

Statistics and MLChecked against a baseline

Forecasting, segmentation, classification and anomaly detection. Every model is compared with a naive baseline, and results that look implausibly good are flagged for review rather than presented as findings.

ModulesSector packages

A package of ready-made analyses for one sector, connected to the organization's data once and run by staff without writing code.

Deployment

One server, departments as groups, each person in their own workspace.

IT assigns a server and the application is installed once. Departments join as user groups and people connect from a browser over the corporate network, wired or wireless, over HTTPS. No cloud tenant is provisioned and no client software is installed.

Reference architecture: sources connect read-only to one server, which serves departments as user groups and is governed by one administrator

CapacityOne server serves the organization; the number of users is set by the licence's seat count.

Departments as user groupsEach group is granted access by category; anything not granted is denied.

One administratorHolds the model keys and the privacy settings; reviews the Knowledge Base; keeps the activity record.

Air-gapped when requiredWith a local model on the server, analysis, dashboards, KPIs and Live Monitor operate without an internet connection.

PlatformsWindows and macOS installers with Python bundled; the macOS package deploys through Jamf Pro.

Smaller deploymentsThe same product runs on a single workstation for an individual professional or a small unit; work created there can be moved to the server unchanged.

Data handling statement

What the platform sends outside the organization, stated in writing. The same terms apply to each supported model, cloud or local.

Datasets, database tables, files and documents remain on the organization's machines and are not uploaded to the model. Databases are read under read-only credentials and cannot be modified.

The model receives a description of the data: its structure and summary statistics and, at default settings, two example rows per table and the most frequent text values. The administrator can switch the example values off for all users. From that description the model writes the analysis, which runs on the server.

Results reach the model only when a user asks for a written explanation, as an aggregated digest the user reads before it is sent, with identifier values withheld. For documents, only the passages the search matched travel with the question.

With a local model, none of the above crosses the network. The server can operate without an internet connection.