R&D product company

AI agents for network management of production

AlFarabi-AI builds AI agents that take over coordination inside a company, give leaders a complete picture of the business and tune production modes towards shared KPIs. The agents run inside your IT perimeter — in oil and gas, metals, business and city management.

From pilot to full-scale deployment — with impact measured against agreed KPIs.

  1. Class 01 A network organisation instead of a pyramid
  2. Class 02 A co-pilot for those in charge
  3. Class 03 Agents that tune production themselves

Product concept

Three classes of AI agents

For centuries companies were run as pyramids because coordinating people was expensive. AI agents make coordination nearly free and change the very shape of management — from the organisation as a whole down to a single well. Al-Farabi described the ideal city as a whole in which every part knows its role; we build that whole out of people and agents.

One management loop

  1. 03Production agents hold operating modes and report reserves and risks
  2. 02The co-pilot turns this into a picture for the leader
  3. 01The network carries the decision to people and agents
  4. KPINetwork KPIs return to the agents as a new target
PYRAMID 6–8 direct reports per manager network KPIs NETWORK OF PEOPLE AND AGENTS Person AI agent Leader co-pilot Production agent

01

Class 01 · Network organisation

From a pyramid to a network of people and agents

For centuries hierarchy rested on the span of control — 6–8 direct reports per manager. The main job of management was coordination: pass a signal up and down, align across, check execution. AI agents take coordination over, and the pyramid gives way to a network where people and agents work as one system.

  • Agents relay signals, reconcile deadlines and track execution
  • People decide, negotiate and own the result
  • Fewer layers — a shorter path from signal to decision

02

Class 02 · Co-pilot for leaders

A private AI advisor for the person responsible for everything

The co-pilot turns scattered data from finance, CRM, HR, correspondence and minutes into a complete picture of the company and helps see the real state of affairs before problems become critical.

  • Closed perimeter: only the leader sees the questions and answers
  • Early risk signals — before they reach a report
  • Every conclusion cites its source: a deal, a document, the minutes
Demo: AlFarabi-AI Core →

03

Class 03 · Self-organising production agents

Production that tunes itself

Agents learn from production data and tune operating modes towards network KPIs — targets built into the company-wide management loop. A well, a furnace or a boiler house works not for a local record but for the result of the whole chain.

  • Each agent knows its own unit and its neighbours' constraints
  • The goal is network KPIs — output, energy, NPV — not one shop's figure
  • A person approves the mode, the agent explains the calculation
Demo: WellHub, the well-pad agent →

Solutions

Four elements

Oil and gas, metals, business and the city each work differently. The method is the same: enterprise data → AI analysis → AI optimisation → measurable impact.

Pumpjack in the Kazakh steppe against the Tian Shan mountains Pumpjack in the Kazakh steppe against the Tian Shan mountains

The element of the subsurface

Production the data sees before the shift does

The history of operating modes, water cut, drilling, failures and network pressure comes together in one loop. The system finds the link between the well, the stock and the control room — and shows where to raise the flow rate, and where not to burn extra energy.

  • Pad mode optimisation (WellHub) — bottom-hole pressures and injection within treatment and injection limits, from daily reports, without a reservoir simulation model
  • Raising the flow rate — quickly find controllable causes of losses and help choose the measures that deliver the greatest expected gain in net production at acceptable risk and economics
  • Well-stock maintenance — repairs by risk, not by calendar
  • Drilling — complications visible in the regime while the well is still being drilled
  • Network energy costs — the optimal mode without guesswork
  • Dispatcher copilot — answers from your own instructions, right at the workstation

Cases and KPIs

WellHub product · result on the demo pad

Well pad operating mode optimisation

Mature field: 85–90% water cut, water treatment and injection at capacity, well modes tuned by hand.

Hybrid proxy models (CRM + LightGBM) trained on daily reports and an optimiser of bottom-hole pressures and injection under facility limits.

  • +10–12%oil within the same treatment and injection limits
  • 5.3%blind 12-month forecast error
  • ≈1 sto compute optimal pad modes

Pilot: 6–8 weeks on one pad WellHub demo →

Target pilot KPIs

Predictive maintenance of ESP wells

ESP failures and unplanned workovers eat production and budget, while repairs follow the calendar.

Failure-risk model on current, load, temperature and repair history; workover queue ranked by expected oil loss.

  • −20–25%ESP failures
  • +15%mean time between repairs
  • 7–14 dayswarning before failure

Pilot: 3 months, 150–300 wells

Target pilot KPIs

Energy efficiency of artificial lift

Power is the largest lifting cost, yet pumps and gathering networks run with a generous margin.

Energy model of wells and network with recommendations on ESP frequencies and pump station modes.

  • −8–12%specific power per tonne lifted
  • 0production lost to optimisation

Pilot: 2 months on one field

Figures are target ranges we commit to in a pilot and verify on your data.

Abu Nasr al-Farabi — two-tone mosaic, the emblem of AlFarabi-AI
Abu Nasr al-Farabi Sage and mathematician from Kazakhstan

Idea

He did not discover new stars. He gave them a language.

Al-Farabi — mathematician, logician and philosopher from the Kazakh land — brought scattered sciences together into a single canon. Logic, geometry, astronomy and music sounded like a score, not noise. He was called the Second Teacher — after Aristotle.

Industry today is where the Otrar library was then: well telemetry, furnace modes, incident logs and shift reports already exist. What is missing is a view that sees the law in them — and a system you can rely on.

Order in knowledge is the beginning of order in business. We assemble operational processes the way al-Farabi assembled the sciences: first a map, then a hypothesis, then a system that keeps working for you.

Path

Fast impact, minimal risk

As in al-Farabi's canon: first the map of sciences, then the proof, then practice. No three-year "AI transformation" without a first number.

  1. 01

    Knowledge map

    Together we find the process with the highest return and frame a hypothesis with an estimate of its financial impact. We start from the data you already have.

  2. 02

    Proof

    We build a working prototype on your time series, not on slide-deck synthetic data. The hypothesis either lives or dies on the numbers.

  3. 03

    Site pilot

    A limited scope, a live shift, a measurable result. You see the impact on the shop floor — and decide whether to scale.

  4. 04

    Canon and handover

    Rollout across the enterprise and growing your team. The capability stays with you: models, procedures, people.

Measure

Systems you can rely on

As in al-Farabi's astronomy: the higher the cost of a decision, the stricter the demands on the method. We don't promise magic. We build a loop that survives the night shift.

Technology

What we build with

Al-Farabi divided knowledge into parts so that it would work as a whole. We combine machine learning, optimisation, LLM agents and computer vision — and connect them to the systems you already run.

Machine learning

Forecasting and diagnostics on your historical series

  • Time-series forecasting: gradient boosting and neural networks
  • Physics + ML hybrids that respect process laws
  • Anomaly detection and early failure warning
  • P10–P90 intervals and factor explanations (SHAP)

Optimisation and scenarios

From a forecast to a decision within constraints

  • Mathematical optimisation with process constraints
  • What-if scenarios and Pareto fronts
  • Decision economics: NPV, cost, risk
  • Back-testing recommendations on history before rollout

LLMs and AI agents

Language that understands your data and documents

  • Multi-agent orchestration: agents analyse, compute and check each other
  • Retrieval over procedures, instructions and logs (RAG) with sources
  • The LLM does not do the maths — it calls models and explains the numbers
  • Commercial and open-source models, including inside your perimeter

Computer vision

Eyes on the line that never get tired

  • Product defects on existing cameras
  • PPE and hazardous-zone monitoring
  • Reading gauges and documents
  • Edge processing next to the line

MLOps and reliability

A model that survives the night shift

  • Model quality and data-drift monitoring
  • Scheduled retraining and version rollback
  • Metrics, logs and a trace for every recommendation
  • Advisor mode before closed-loop control

Security and perimeter

Your data stays with you

  • On-premise or private cloud in Kazakhstan
  • No data leaves your network
  • Role-based access and audit log
  • Compliance with personal-data requirements

Integration with corporate systems

We connect to what already works — no system replacement, no manual exports.

  • ERP and accountingSAP, 1C:ERP, Oracle EBS
  • OperationsSCADA / DCS, MES, data historians (AVEVA PI)
  • Maintenance and assetsSAP PM, IBM Maximo, 1C:EAM
  • AnalyticsPower BI, Visiology, DataLens, Superset
  • CityGIS, transport GPS tracking, smart metering, request systems
  • DocumentsEDMS, wikis and knowledge bases, email, Telegram bots
  • ProtocolsREST API, OPC UA, Modbus, Kafka, Excel / CSV

Contact

Let's look at your processes together

Tell us where it hurts: the well stock, a furnace, the reporting, a route or the heat network. We reply within one business day. The first meeting is about data and impact, not slides.