Andrey Melkov

Speed attracts attention.

In manufacturing, it's increasingly a process story.

One factory spends hours searching documentation.

Another asks an AI assistant and gets the answer in seconds.

One factory discovers a machine failure after production stops.

Another sees it coming days in advance.

One factory relies on manual inspection.

Another has cameras checking every product around the clock.

The difference isn't intelligence.

It's friction.

AI keeps removing small pieces of friction from industrial systems.

It happened when companies stopped treating AI like a science project.

For years, everyone was running pilots, demos, and proof-of-concepts. Most of them never went anywhere.

Turns out the hard part wasn't building the AI.

The hard part was cleaning up years of messy data, integrating systems, and figuring out where AI could actually make a difference.

The most successful use cases aren't flashy.

They're things like predicting machine failures, helping engineers find answers faster, automating inspections, and reducing downtime.

That's probably the most important shift happening right now:

People are asking less "What can AI do?"

Hybrid is best: buttons for clear paths, AI for messy questions.

Connect it to your CRM. Otherwise it's just a fancy FAQ.

Measure: how often does the bot solve it alone? How fast? When does it fail?

Pilot for a month. Watch what breaks. Fix. Add more.

Delegation is not the same as having a system.

I left my business for a year. Thought it would run itself. Came back to zero.

Why don't owners build processes? Not because it's hard. Because they're scared.

Scared it won't work. Scared employees will quit. Scared someone smarter will join. Scared to see the truth.

They thought mechanics were slow. Nope.

Mechanics were WAITING.

- Writing in paper logs

- Waiting for a supervisor to walk over

- Waiting for parts room to get the message

So we automated the handoffs.

Now a mechanic uses a mobile device. Finish one step – next step opens automatically. No paper. No running around.

Find a defect? The system tells the supervisor AND the parts room instantly. The warehouse reserves parts before the mechanic finishes the inspection.

Result: 150 minutes. Six trucks a day. No new hires. No new lifts.

The process just… works now.

Moral: don't blame people. Look at where they wait.

Bad approach.

Better approach: a team of three AI agents.

- Navigator – finds the exact 30‑second window where Ivan spoke.

- Eyes – looks at that window and sees "Ivan, slide with 'delivery: May 15'".

- Master – checks if that's enough, then answers: "At 23:45, Ivan said May 15."

Accuracy goes from 64% to 75%. Not magic – just not throwing away information.

How to build it (5 steps):

1. Get subtitles with timestamps (Whisper works).

2. Split into chunks, keep timestamps, put in a database.

3. Teach Navigator to find relevant time windows.

4. Add Eyes only for visual questions (slides, objects, people).

5. Master collects everything and gives a timestamped answer.

Rule #1: never lose timestamps. Rule #2: never analyse the whole video at once.

The audience wasn't huge – but it was the RIGHT people. Technologists. Production heads. Owners. People who actually do the work.

What I liked: talking to people who get it. That's rare and valuable.

What I'll change next time:

1. One hour is a joke. We need 1.5‑2 hours. Deep dives take time.

2. I want to show one complete process from start to finish (blueprint → tech card → price). But then I can't cover everything. Maybe a series. One webinar = one deep story.

3. More live questions. When the audience is silent, I don't know what they need. Next time: ask questions in advance, or bribe people to unmute.

So – if you work in industry or with AI – what would actually help you?

- Case studies?

- Live demo?

- ROI math?

- Our screw‑ups (there are many)?

Here's what neuroscience says: the brain predicts, does something, sees what happened, compares, fixes. LLMs don't do that. They just generate text.

We need three things AI is missing:

1. **Real actions** – not "I'll write that I sent an invoice" but actually calling the API to create it, then checking it exists.

2. **Hierarchy** – the brain stores goals and rules at the top, messy details at the bottom. AI loses important stuff in the middle of a long chat. So store structure (deadline, budget, agreements) separately from chat noise.

3. **Memory that writes back** – not just reading old docs (RAG). The AI needs to remember what went wrong last time, then at night "sleep" on it and turn mistakes into rules.

Also – **invariants**. Hard rules that cannot be broken. Like: no invoice without bank details. No discount above 10% without manager approval. Check these AFTER every action but BEFORE replying to the customer.

Five steps to make it work:

1. Turn each deal into a structured card (15‑30 fields)

2. List allowed actions (only these, nothing else)

3. Write down 10 invariants

4. Log everything

5. Nightly processing

Panic. "See? AI is useless."

But the study said: 95% of companies never scaled past pilot. They tried a little, didn't commit, gave up.

That's not AI's fault. That's bad implementation.

Also, Andrej Karpathy: chatbots beat the average human. But not the expert human. Your best lawyer is still better than ChatGPT. For now.

So AI isn't a flop. It's just not magic. It's a tool. Some people use it well. Most don't.

If you want it to work: clean your data, train your people, start small, and don't expect miracles.

Why? Because its brain (the knowledge base) never changed. The world changed, but the AI didn't.

You think: "Let's make it save its own good answers." Good idea. But if you save everything, you also save its mistakes. Then it learns from its own lies.

We built a filter. Before any answer gets saved, three checks:

1. Can you prove it with a real document?

2. Is the link real?

3. Is this new information (not a duplicate)?

Result: The AI learns slowly but never gets dumber. It adds only good stuff.

There's a Russian service. You pay in rubles. No VPN. They give you access to ChatGPT, Midjourney, DeepSeek, all of them.

Also they have a drag‑and‑drop automation thing (n8n). You can connect email → AI → Telegram in 5 minutes. Costs about 22 rubles a day when you use it.

Data stays in Russia. No foreign servers.

First client: a factory that had been trying to implement 1C for two years. Nothing worked. People hated the system.

We came in. Four months. We didn't add anything fancy – just cleaned up the mess. Found €200k just sitting there in bad logistics, double entries, unaccounted stock.

Then a sausage company. Farm to fork. That took 1.5 years.

After that we thought: we need a logo with meaning. So we used runes. ᛏ for determination, ᚨ for communication, ᛒ for growth.

We put an AI on a local server (no cloud because state secrets). It reads scans. It types into 1C. 98% accuracy. 10 seconds per doc.

Now 7‑8 people check the AI's work and handle weird cases. The rest moved to real accounting. Month‑end doesn't require panic anymore.

Also saved 900k rubles a year on outsourcing.

We built an AI that reads the PDFs. It finds the INN, the OGRN, the cadastral number, the date, the amount. In 30 seconds.

Accuracy 90%. Errors down 80%. The humans now check the AI's work instead of typing everything.

They don't hate their job anymore. The company doesn't get sued anymore.

That was the factory.

Now a camera watches every sheet. A neural network that was trained on 10,000 defect photos flags bad ones in real time.

The human just sits and watches the AI. Occasionally says "yes" or "no" on tricky cases.

Undetected defects down 70%. Workers don't go home with headaches.

AI didn't replace them. It saved them from a mind‑numbing job.