Home → Stop the leaks → Automation and AI
Stop the leaks · direction 02

Automation that pays back within a quarter

Where a human relay of information used to sit, an agent takes over. Models run on premise, inside your own perimeter — no cloud, no data leaving the building. We write the software ourselves, so we own the result end to end.

Sound familiar?

Six signs routine is eating the company

01

People spend half the day moving information from one window to another.

02

Requests arrive round the clock but are only processed during office hours.

03

The manager learns about a deviation from a report at the end of the month.

04

Reporting is assembled by hand in spreadsheets and is out of date by the time it is read.

05

Nobody actually knows how loaded the equipment and the people are.

06

A competitor automated and is taking your clients on speed alone.

What we do

Five things we actually build

Each is bought separately and pays back on its own. Together they form a loop that holds the result without you.

AI agents on routineReceiving and routing requests, assigning work, accepting completed jobs by photo, reporting
Executive dashboardA weekly summary with losses in rubles and instant alerts on deviations
Apps for your processRounds, photo-verified checklists, work orders, standards in the employee's phone
Computer visionUtilisation of people and equipment from cameras, quality control, safety compliance
Integrations1C, CRM, telephony, warehouse — data stops living in five separate spreadsheets
On-premise modelsEverything inside your perimeter, no cloud and no external subscriptions
Zhilishchnik case: an AI agent listens to the working chats in Telegram and MAX, creates requests itself, notifies the crews and hands them to the manager to close — freeing more than half of the dispatcher's time. There was no resistance: people stayed in the channels they already used.
Results

Measurable and verifiable

−60%dispatcher time · Zhilishchnik
24/7request intake without a human
2–4 wksto a working pilot
on-premisemodels inside your perimeter
For the technical buyer

Under the hood

This section is for the IT director and the security team who will review the solution after you. The stack without marketing: what we use, why, and what it gives the business.

Where the agent gets its knowledge
RAGretrieval-augmented generation

The agent answers from your documents — procedures, contracts, ticket history — not from the model's own memory. It retrieves the relevant fragments first, then answers from them.

Every answer is verifiable: it points to a specific clause in your document.
Vector indexembeddings and chunking

Documents are split into fragments and converted into numeric vectors. Retrieval works by meaning, not by keyword match.

A question phrased in the caller's words still finds the right clause, even when the wording differs.
RAG or fine-tuningwhen each applies

Nine tasks out of ten are covered by RAG: cheaper, and updated in a minute when a procedure changes. Fine-tuning is used only for narrow terminology and tone.

You don't pay to retrain a model every time an instruction changes.
What it runs on
Local LLMson-premise, air-gapped

The model is deployed on your own hardware, inside the perimeter. No request leaves it — internet access is not required to operate.

The only option that clears a security review at a bank or a restricted-access facility.
QuantisationINT8 / INT4

The model is compressed to fit the GPU you actually have, with controlled quality loss on your specific tasks.

The difference between a single server and a rack, for the same result.
Context windowtokens

How much the model holds in one pass. The limit is handled by putting only relevant material into the window, not by buying a bigger model.

Works across tens of thousands of documents, not just a short instruction.
How the agent acts, not just talks
Harness layeragent orchestration

On its own a model only generates text. The harness gives it tools, memory, a self-check loop and hard limits on steps and cost.

Without it an agent can't file a ticket and won't stop when the data runs out.
Function callingcalling your systems

The agent invokes functions: create a deal in the CRM, write off stock in the ERP, fire a webhook, run a SQL query against your database.

The output is a record in your system, not a message in a chat.
Voice pipelineSTT → LLM → TTS, VAD, SIP

Speech recognition, response, speech synthesis. VAD detects when the caller has finished. Connected to telephony over a SIP trunk or softphone.

Outbound and inbound calls run on the phone system you already have.
Why it can face customers
Guardrailscitations and refusal

The agent must cite a source and must say it doesn't know rather than invent. Topics outside the procedure are blocked at the entry point.

An invented answer to a customer costs more than an unanswered question.
Evalsregression tests for quality

A set of real questions with reference answers, run after every change to the prompt or the knowledge base.

A fix for one case doesn't quietly break a hundred others — you see it before release.
Audit logand PII redaction

Every request, every tool call and every decision is written down. Personal data is stripped before it reaches the model.

There is something to show an auditor and something to reconstruct a disputed conversation from.

On air-gapped environments specifically. We have worked with sensitive data in isolated networks where internet access simply does not exist. In that configuration the model, the vector store, the harness and the logs all live on your hardware, and updates arrive through a controlled gateway.

Where the perimeter is mixed, personal data is redacted before it reaches the model, and the data route itself is documented in a form you can attach to your data-protection policy.

Questions

Frequently asked

Will our data go anywhere outside?
No. Models are deployed locally, inside your perimeter — no cloud and no external subscriptions. The same applies to video analytics: the server sits with you and the stream never leaves the site.
Our equipment is old, there is nothing to connect to
Age does not matter. Where there is no CNC or sensors we count from cameras: the system sees the work zone, the operator and whether work is happening. A machine from the eighties is monitored the same way as a new machining centre.
We bought a system before and nobody uses it
That happens when software is installed without touching the process. We go the other way round: first we work out how the job is really done, rebuild it, and only then lock it into software. Often what you bought can be revived rather than replaced.
What stays when you leave?
A working system, documentation, source code and a trained owner on your side. We do not keep clients on a leash: the goal is that after a year you can choose not to renew and lose nothing.
Why is RAG better than just uploading documents to a chatbot?
Because the knowledge base lives separately from the model. When a procedure changes you update one document and the agent answers from the new revision the same minute — no retraining, no prompt rewriting. Every answer also carries a reference to the source fragment, so a disputed case is settled in a minute instead of becoming an argument about what the model supposedly said.
Our network is air-gapped. Is this even possible?
Yes, and we have done exactly that. The model is deployed locally on your server; the vector store and the logs stay inside the perimeter. No external service takes part in operation — internet access is needed only to pull the initial images, and only through a controlled gateway. Your security team gets a data-flow diagram and access to the audit log.
What hardware does this require?
It depends on the task, but the numbers are far less alarming than people assume. A text agent over a document base serving tens of users runs on a single consumer-class GPU: quantisation compresses the model without noticeable quality loss on a narrow domain task. The exact configuration is sized during the assessment, based on load and response-time requirements.
HEIM Club

Your case reviewed among peers

A private club for business owners: live offline case reviews in central Moscow, a table assembled around your request, business games and access to a team of practitioners. Everything discussed stays inside the circle.