Does Private AI really work without an internet connection?

Yes. Speech recognition, speaker diarization and the language model all run on local GPUs from weights stored inside your network. Nothing in the pipeline needs an outbound connection.

Which languages does meeting transcription handle?

Ukrainian, Russian and English, including meetings where people switch between languages mid-sentence — the speech model was chosen for exactly that.

What hardware do we need?

It depends on volume. A single team's meetings fit on modest GPU capacity; scaling to the whole company means adding GPUs, not redesigning the system. We size it from your numbers during the requirements phase.

Can tasks go into the task manager we already use?

Yes — through its API or a webhook. Until that integration is agreed, minutes and tasks are available in an internal dashboard.

Is a local model accurate enough for forecasting?

In our forecasting project for a food manufacturer the in-house model matched actual sales at 80–90%, on a quarterly horizon with weekly re-forecasts.

Private AI

AI that never leaves your perimeter.

Meeting minutes, task assignment and demand forecasting on models that run on your own servers — no cloud AI services, no public APIs, no data leaving the network.

Your servers, isolated network

Where it runs

None — no public AI APIs

External calls

Ukrainian, Russian, English, mixed

Languages

80–90% forecast match

Proven result

Why private, not cloud

The data that would make AI most useful — sales plans, prices, what was decided in a management meeting — is exactly the data most companies cannot send to a cloud model. Legal, security and sometimes plain common sense say no.

So we build the other way round: open-weight models deployed on the client's own hardware, inside a network segment with no route to the internet. Recordings, transcripts, forecasts and model weights all stay on premises, and every stage of the pipeline is something the client's IT team can inspect.

Where it already runs

Two projects, both under NDA — described without names.

Agriculture · in delivery

Meeting Intelligence for a diversified agroholding

Project in delivery, starting with a proof of concept

Meeting recordings transcribed with speaker diarization, so the text shows who said what.
A local LLM turns the transcript into minutes: topics, decisions, open questions.
Agreed actions become tasks, matched to owners from the speaker labels and the attendee list, and sent to the company's own task manager.
Fully closed perimeter: an isolated VLAN with no outbound internet route; no stage calls an external API.

Manufacturing · delivered

Demand forecasting for a large food manufacturer

80–90%forecast matched actual sales

Quarterly sales forecast with weekly re-forecasts as actuals come in.
Built on a local model — no public APIs, so commercial data never left the company.

How Meeting Intelligence works

A five-stage pipeline, every stage on the client's own hardware.

Intake

An audio or video recording arrives through a watched folder or a small internal web UI, with the list of attendees.

Transcription and diarization

Speech recognition (Whisper large-v3, offline) with speaker separation (pyannote); the speech model was chosen for its accuracy on Ukrainian, Russian and mixed speech.

Minutes

A local open-weight LLM turns the diarized transcript into structured minutes: topics, decisions, open questions.

Tasks

The same model extracts agreements and action items, assigns owners from speaker labels and the attendee list, and structures each task.

Delivery

Minutes and tasks land in an internal database and dashboard, and go to the existing task manager through its API or a webhook.

Intake

An audio or video recording arrives through a watched folder or a small internal web UI, with the list of attendees.

Transcription and diarization

Speech recognition (Whisper large-v3, offline) with speaker separation (pyannote); the speech model was chosen for its accuracy on Ukrainian, Russian and mixed speech.

Minutes

A local open-weight LLM turns the diarized transcript into structured minutes: topics, decisions, open questions.

Tasks

The same model extracts agreements and action items, assigns owners from speaker labels and the attendee list, and structures each task.

Delivery

Minutes and tasks land in an internal database and dashboard, and go to the existing task manager through its API or a webhook.

What "closed perimeter" means in practice

Deployed in an isolated network segment with no outbound route to the internet.
Model weights, recordings and transcripts are stored on the client's own storage, with retention rules the client sets.
No cloud AI service and no public API anywhere in the pipeline.
An audit log of every recording, transcript, summary and task.
Sized for the first team's volume and scaled by adding GPU capacity, without re-architecting.

How a Private AI project runs

Small first, on real data, before anyone buys hardware for the whole company.

Proof of concept

The pipeline on a handful of real recordings or a slice of sales history, so quality is judged on your data, not a demo.

Requirements and target architecture

Volumes, languages, integrations and security constraints written down; hardware sized from them.

Infrastructure

GPU servers, storage and the isolated network segment, built with your IT team.

Build and integration

The pipeline, the dashboard and the integration with your task manager or planning systems.

Testing, rollout and handover

Acceptance on agreed scenarios, rollout to the first team, documentation and support.

Frequently asked questions

Where Private AI fits

Agriculture

Agroholdings: meeting minutes, planning data and documents that stay in-house.

Learn more →

Manufacturing

Demand forecasting and plant knowledge on models that run on your servers.

Learn more →

AI services

Assessment, use-case selection and the controls that make AI safe to run.

Learn more →

Start with a proof of concept on your own data

Tell us which meetings or which forecast you want to automate — we'll scope a pilot that runs entirely inside your network.