Bus and car collide on Severnaya, one lane closed
Traffic towards the centre is slow.
✓ Published · confirmed
AI agents
People set the rules. Agents do the work.
the agent thinks for you
Igor Karmazin · AI Technical Product Operations
I build AI agents and systems you can trust with real work.
My AI agents run city news channels on their own - from hundreds of sources to a finished post. EnergyAI Ops turns equipment readings into clear decisions for engineers. Behind both: 9+ years in power engineering, where mistakes are expensive.
Experience and platforms
- MTS
- NESK Electric Networks
- WAIC 2026
- Telegram
- MAX
- VK
Hundreds of sources in. One post out. And a reason behind every decision.
Sound familiar? These are not made-up problems. I have taken each one apart by hand - in my own channels and on real sites.
- Media
The best story ran on a competitor's channel
An admin scrolls through hundreds of sources by hand. By the time they decide what matters, the city has already read it elsewhere.
- AI copy
The model writes - and invents
You ask it to rewrite a story, and it adds ages, job titles and "details" that were never in the source.
- Publishing
"Sent" doesn't mean "published"
The script reports success, but the channel is empty. Or it never got a reply and posted the same story twice.
- Engineering
Plenty of data, no picture
Readings live in SCADA, repairs in a logbook, manuals in a cabinet. The cause of a deviation is hunted down from memory and phone calls.
One principle in every system
- 01
Agents do the work
They select, compare, verify and write. Each has its own role and criteria - just like people in a newsroom.
- 02
People set the rules
Boundaries, tone and judgement calls stay with the people accountable for the outcome. AI speeds things up; it does not give orders.
- 03
Results live in the real world
Every decision is logged with its reason. Success is a published post and a verified result - not a green tick in a report.
System 01 · MediaLive
A newsroom run by AI agents
Each of our city channels is run by a team of AI agents built like a real newsroom. Some pick the stories, others catch repeats and check facts, others write the copy and stay on call for breaking events. People set the rules. Every agent decision is logged with its reason.
How a story reaches the feed
- Step 01
Collection
Around the clock, the system listens to hundreds of city and regional sources: Telegram channels, websites, agencies and emergency services.
- Step 02
Central filter
Ads, digests, empty posts, job notices and the same story from ten different sources are filtered out before an editor ever sees them.
- Step 03
Agent editorial meeting
On a schedule, editor agents weigh the strength of each event and pick one story - the one that matters most to the city.
- Step 04
Rewriting
The story is retold in original words for the people of that city - no clickbait, no guesswork.
- Step 05
Publishing and verification
The post goes out, and an observer bot confirms it actually appeared. No confirmation means no resend - and no duplicates.
What happens inside? A trade secret I show the result, not the recipe. Hundreds of stories in. A closed newsroom inside. One post out - and it actually went live.
In · hundreds of sources
Website and social media parsers
- Websites
- Telegram
- MAX
- VK
- Agencies
AI watches photos and video
Agent newsroom
ClassifiedOut · one post
Decision dossier
- Agent 01recommend
- Agent 02nominates
- Agent 03reject
- Vision agentmedia checked
- Chief editorselected
Illustrative example
Before → after
The usual wayWith the agent newsroom
SelectionA person scrolls the feed and picks by eyeAgents select against the city passport; people set the rules
CopyRewrites with made-up details and clickbaitOnly facts from the source, in the voice of each city
RepeatsOne story from five sources becomes five postsOne post - developments go out as replies to it
Publishing"Sent" and hoping it went outA bot confirms delivery; no duplicates
Newsroom rules
- 01
Only facts from the source.
- 02
One story, one post.
- 03
Silence beats a weak post.
- 04
At night, only what's urgent.
- 05
Every city has its own passport and its own voice.
- 06
People write the rules. Every decision is logged, with its reason.
System 02 · EnergyConcept and prototype
EnergyAI Ops
Equipment, energy and maintenance - in one engineer's workspace
It plugs into the meters, sensors, SCADA and spreadsheets a site already has. It builds a digital equipment passport, spots deviations, explains their likely causes and checks whether a repair actually worked.
EnergyAI Ops does not replace SCADA or BMS - it adds what they lack: shared context, analytics and a clear engineering process.
How the system works
- Step 01
Connect the data
EnergyAI Ops connects to the site's existing infrastructure. No equipment needs to be replaced.
- electricity meters
- sensors and controllers
- SCADA, BMS, DCIM, EMS
- Modbus RTU/TCP, RS-485
- APIs, CSV and Excel
- maintenance logs
- Step 02
Digital equipment passport
Manuals, procedures, specifications and maintenance history become a structured digital passport.
- type, model, serial number
- specifications and operating limits
- normal modes
- connected meters and channels
- event, maintenance and repair history
- documentation and engineers' notes
AI helps extract information from documents, but safety-critical fields are approved by a specialist.
- Step 03
Observation and baseline
In the first days, the system learns how the equipment behaves and checks data quality. Initial observation usually takes 2 to 5 days. At this stage EnergyAI Ops gives no final diagnosis - it builds an initial baseline, a model of normal equipment behaviour.
- the site's operating mode
- data quality and completeness
- seasonality
- number of operating cycles
- emergency and rare modes
Model maturity depends on these factors.
- Step 04
Detect deviations
The platform compares current readings with the normal mode and flags unusual changes.
- rising energy use
- phase imbalance
- falling power factor
- unstable operation
- running outside the schedule
- changes after a repair
- Step 05
Analyse the context
EnergyAI Ops does not treat every signal as a failure. It checks the context - which cuts useless alerts and makes recommendations better grounded.
- production schedule
- temperature and weather
- load changes
- recent maintenance and switching
- neighbouring equipment
- history of similar events
- Step 06
Engineer confirmation
The engineer remains the decision owner. Feedback is stored and refines models, rules and baselines.
- normal mode
- needs observation
- needs inspection
- technical action required
- false alarm
- confirmed fault
- Step 07
Verify the result
After maintenance, EnergyAI Ops compares readings before and after the work. Maintenance becomes a closed, verifiable loop.
- did the deviation disappear
- did the reading return to normal
- did energy use change
- does the problem persist
- is another check needed
An anomaly is a signal plus context A spike on a chart is not a failure. Press “Check the context” and see what the system finds out before it calls an engineer.
Ventilation unit · Building B
Energy use above baseline at comparable load
- Production schedule-
- Temperature and weather-
- Current load-
- Recent maintenance-
Engineer Copilot
…
Engineer's decision
Illustrative example
Five agents, each with its own job A Supervisor / Orchestrator coordinates them. Every engineer confirmation flows back through a Feedback & Model Governance loop and refines the models.
Supervisor / Orchestrator · Feedback & Model Governance
Data Guardian
Monitors the quality of incoming data.
- gaps
- errors
- time synchronisation
- invalid values
- source stability
Passport AI
Works with documents and digital equipment passports.
- extracts specifications
- structures manuals
- links documents to assets
- helps find the right information
Baseline AI
Builds the model of normal operation.
- operating cycles
- typical load
- seasonal changes
- acceptable ranges
- stable operating patterns
Context Analyst
Analyses a deviation together with its context.
- production schedule
- weather
- repairs
- load
- related events
- equipment history
Engineer Copilot
Explains the situation in plain engineering language.
- what changed
- why the event needs attention
- which causes to check
- which data supports the conclusion
- which actions to consider
System 03 · ProductLive
Voltovich
A personal assistant for anyone connecting a plot or a house to the grid. It runs in MAX and on voltovich.ru: it tells you where to apply, explains every field of the application, breaks down the contract, technical conditions or a refusal, and reminds you of deadlines.
A product built from domain expertise: years in the grid business, packaged into an assistant and a content pipeline.
voltovich.ru ↗A route for your property
Not a generic manual, but your next step - based on your region, your property and the grid company's reply.
Calculators and a knowledge base
Connection fees, required capacity, plain-language guides, tariff news and outage updates.
Honest boundaries
It never files on the user's behalf, never asks for passwords or codes, and never makes things up.
How I work No six-month spec, no big promises. First we prove it works, then we scale.
- 01
Review
I look at the process as it really is: where time is lost, where errors happen, where AI will genuinely help - and where it won't.
- 02
Pilot
I launch on one channel, one city or one site - in test mode, alongside current work. Nothing gets broken.
- 03
Scale
Once the result is confirmed in the real world, I roll it out further - one step at a time, with a way back.
An engineer who builds agents
I came to AI not from slide decks but from switchrooms, generator rooms and server rooms. There, a system either holds under load or it doesn't - there is no third option. I build agents the same way: they have to work, not just look good in a demo.
- Kuban State Agrarian University, 2015 - Engineer, electrification and automation
- Maykop State Humanitarian and Technical College - Software for computer and automated systems
- 2025 -
EnergyAI Ops · Founder
Concept and prototype of an AI platform for engineering operations: equipment passport, agents, maintenance verification.
- 2021 -
MTS · Critical Infrastructure Engineer
Power for telecom infrastructure: UPS, diesel generators, server rooms, commissioning and contractor supervision.
- 2018 - 2019
Risberry · Management Director
Production planning, quality, safety and equipment maintenance.
- 2016 - 2018
NESK Electric Networks · Electrical Engineer
Grid connection, metering, single-line diagrams and investigations of unmetered consumption.
- 2014 - 2015
Dnepr-Gold · Site Electrician
Diesel generators and industrial equipment at a remote site.
Recognition
WAIC Future Tech OPC Challenge 2026
Online selection, the Xi'an regional semifinal and a place among the top 160 projects in China.
Karmazin.Media - First place
An AI project automating marketing, content and business processes in an entrepreneurship program.
AI News
What matters in AI - briefly, with a primary source and a practitioner's view.
This section is in preparation. It will feature concise analysis of meaningful changes in AI, with verifiable primary sources.
All articles
stand out from the crowd ↘
A newsroom, a content pipeline, engineering data or an EnergyAI Ops pilot at your site. Tell me the problem - we'll work out where AI delivers and where it doesn't.