How I use AI in GymBeam

0:00Everybody talk about it. This is real

0:03AI. Human is battle neck of [music] a

0:06system in an AI. All these record cost

0:08more than €200,000.

0:10This is better way of buying. It's much

0:13easier to ask for forgiveness than for

0:16permission.

0:477 a.m. I finished my tennis class. The

0:51most of cos start their day with the

0:53mail inbox but I used to start my day

0:56here. It's something like a biohacking

0:58for me. When we talk about the

0:59biohacking the most of people think

1:01about red light therapy, NAD plus cold

1:05plunges. But I think that this is better

1:07way of buying that in an early morning

1:11have a sun in my eyes, fresh air,

1:14nature, fun with my friends and of

1:17course sweat with the tennis. And after

1:20this, I used to have a cold shower and

1:23it's something like a dopamine spike and

1:25a great start of the day. But today will

1:28be totally different. Today will be

1:30about the AI. I will show you use cases

1:33from my company. How we built

1:36competitive advantage, how we implement

1:39AI to processes in our company. And I

1:42think that it can be inspiration for

1:44your company and your career. How to

1:46build modern techsavvy company.

2:04We just arrived to our offices in

2:06Akosita, Slovakia. And here on this

2:08address work about 400 my colleagues.

2:11And here we have a big building with

2:13about 20,000 square meters. And it's not

2:16only about the offices here. We have

2:18also manufacturer for a few of our

2:20products. Another production facility is

2:23in a Germany and we have offices in the

2:2515 cities. But this is the biggest and

2:27many people think that I'm a crazy about

2:30reporting during my doctoral studies was

2:32my main topic business intelligence in

2:35e-commerses and I create so many catchy

2:38reports like this. And this reporting is

2:40real time. It's easy to check few KPIs

2:43and in the different offices we have

2:45different reports that in this building

2:47is more than 20 display like this but

2:50many people think that I'm crazy about

2:52business intelligence and reporting and

2:54it is true cuz we have a here much

2:58bigger report and in our company we have

3:02few technologies for visualization of

3:04this but this is in HTML our own

3:07visualization of our datas it's very

3:10easy to manage manage reports like this

3:12with the AI agents. And today will be

3:14main topic of this video AI and AI

3:17agents. And one of the from my point of

3:20view very useful use cases is that we

3:23have on our Slack AI agents and one of

3:26them Bob is responsible for reporting

3:28and I can write to him a prompt. For

3:31example, what is this number? How we

3:33calculate it? And he respond me in a few

3:35seconds how we calculate this number.

3:37And also I can write to him please

3:39change methodology behind this number

3:41that for example don't count B2B orders

3:44in this number and he can change it in a

3:47few seconds and it's very useful because

3:49we can do this 24 hours 7 days in a week

3:52and I can change this report in Sunday

3:54evening in a few seconds AI agents help

3:57us so much but how look like our digital

4:00colleagues our AI agents I can show you

4:02in our AI server

4:06here is a special room for uh servers

4:09and we have a here very traditional

4:12server for internet in this facility

4:15work hundreds of our colleagues and they

4:17use uh more than thousand devices and we

4:20have a three servers like this but we

4:22are very proud on this this is our AI

4:25server and here is our AI autonomous

4:28agents here we have a Bob and Bob is

4:31very special autonomous AI agent who is

4:34responsible for reporting that everybody

4:37from company can write on a slack a

4:39message to Bob and ask him about some

4:42datas and this agent can share datas

4:45with our colleagues but also this agent

4:48can create reporting for our colleagues

4:51and it's not necessary only in a tableau

4:54which is our main visualization layer

4:56but also this agent can develop

4:59reporting it HTML in a programming

5:02language I think that it's totally

5:04gamecher that we can have a datas very

5:06fast with a very good accuracy and we

5:09can develop very complex reporting in a

5:12few minutes and here is our another AI

5:15agents for example this uh agent Ava

5:17it's my personal assistant and

5:20interesting thing is that for example

5:22this agent Ava can create reporting too

5:25and have a more datas than Bob for

5:27example datas about salaries in a

5:29company is not shared with the Bob but

5:32it's shared with the AVA and access to

5:34AVA have only me and few colleagues and

5:37it helped me so much with the reporting

5:39want to calculate rentability and in our

5:42datas I need a use salaries in our

5:45company but it is not so good if I share

5:48with a Bob how big salary is somebody in

5:51a company another interesting agents are

5:54for example Romy it's for our finance

5:56department for accounting or Shelly is

5:59for our ager department it's something

6:01like a personal assistant who help

6:03communicate with the candidates It's

6:05book meeting room for interview send

6:07some email to our candidates and this is

6:10another important agent it's Mia Mia

6:13it's our specialist on account creation

6:16I'm uh very proud that this agent can

6:19create a product detail page in our

6:22eShop that if you see so many datas in

6:24our e-commerce system there is row datas

6:27and this Mia agent create from these row

6:30datas product detail informations like

6:32long description of a product and it's

6:35not only about description. I think that

6:37interesting part is that we have a logic

6:40how much tests of our products we need

6:42and this MIA can create order in our

6:45system and send products to laboratory

6:48for tests and after that uh laboratory

6:51send us results. This MIA can upload

6:54results of tests of our products on a

6:56product detail page that for many people

6:58it can be surprised that in our product

7:01detail page in a few products we have

7:03more than 20 tests of ingredients and

7:06products and way how we create this test

7:08is that this agent create order in our

7:11system send these products to laboratory

7:14tests. after laboratory send to this

7:17agent results. This agent upload these

7:20results on our website. Another very

7:23practical use case of our agents is

7:25Eric. Eric is our operational purchasing

7:28specialist. What does it mean? We have a

7:30few warehouses in Europe and we have a

7:33few production facilities and thousands

7:36of subscribers. and Eric have a forecast

7:39uh how much products we will need in uh

7:41which warehouse when and Eric sent

7:44orders to production facilities for

7:47ingredients to our products. He sent

7:49orders to packaging materials and send

7:52orders to logistic companies to order

7:54transfer from manufacturer one to

7:57manufacturer two and this bot can send a

8:00thousand of orders to thousands of a

8:02manufacturer very effective way with a

8:04great results. It's very practical use

8:06case for a agents because this agent

8:09calculate so much numbers on the back

8:11end of our company and send the orders

8:14to so many another companies and manage

8:17material flow and supply chain in

8:20thousands of venues in Europe. Another

8:23practical use case is Samir. Samir is

8:26our logistic expert. He monitor so much

8:29datas in our logistic that we have so

8:31many sensor on a can wear systems in our

8:34logistic and collect the datas about

8:36that and uh can predict some issues if

8:38some issue happened after that can

8:40communicate with our colleagues on many

8:43departments what's happened and what we

8:45should do and it's work 24 hours 7 days

8:48in a week that it's very fast delivery

8:50of information to our logistic experts

8:53and engineers what we should solute and

8:56another interesting agent is a VEX. VEX

8:59is our payroll specialist that we have a

9:02employees in about 20 countries where we

9:05can calculate salaries in a 20

9:07jurisdictions and every country have a

9:11different system of calculating

9:13salaries, benefits and things like this.

9:15And it helps us so much with the

9:17information sharing with our uh payroll

9:20department but also communicate with our

9:23employees cuz uh so many colleagues from

9:25these countries want to write messages

9:28to our colleagues but it's more

9:30effective if they write messages to a

9:32agent and he respond so many requests

9:35from them but few requests recent on a

9:37human if it's need some offline

9:40interaction or help with something it

9:42save hundreds of hours of our ag

9:44department. Another devices in this

9:47server is Nvidia DGX Sparks and we have

9:51a three devices here. The first is for

9:54translations and uh now we have a local

9:57model for translation. It's from my

9:59point of view it's so effective because

10:02about two years ago we had 42

10:05translators in our company and they

10:07translate usually from English to local

10:10language and in two years we cut this

10:13head count on a zero and nowadays we

10:16have a zero translators in our company

10:18and this is equivalent more than 42

10:21translators in our company because two

10:23years ago we don't have so many markets

10:25for example last half year we launched

10:29France and Netherland market and it was

10:31100% translated with a local model on

10:34this DG Xpark and here is to another

10:38with the general intelligence and we use

10:40it for very specific use cases but the

10:43most interesting part of this server

10:45from my point of view it's our Nvidia

10:48H200 it's here in this panel and it's

10:52very expensive graphic card but it's big

10:55proud for us that we have a great

10:58general general intelligence model which

11:00run on this hardware and hundreds of our

11:03colleagues can prompt them and have a

11:05respond very fast. The first impressions

11:08from our own model is that many people

11:10have experience with the open AI or

11:13entropic and they have experienced that

11:15their model responded in a maybe 15

11:18seconds or 20 seconds and our own model

11:21responded in less than 3 seconds that

11:23it's totally different uh user

11:25experience and our colleagues love this

11:28and here we can see how it work here is

11:30many Mac minis it's very cheap device it

11:34costs maybe $1,400 €100 and we have a

11:39more than 20 magnes like this. And here

11:41is our autonomous AI agent wax

11:44responsible for salaries. This small

11:46device can replace few full-time

11:49equivalent jobs on our ager department.

11:51But few of these devices I think that

11:53can replace for example this model for

11:55translations. It can replace about 42

11:58full-time equivalent employees. And I

12:00think that in this rack we have

12:02equivalent of hundreds of our employees

12:05and few of these agents don't only help

12:08like replace full-time equivalent job

12:10but help to hundreds of our colleagues

12:13with a small part of their job. For

12:15example like a Bob who is responsible

12:17for reporting.

12:28Okay. And here is real AI. Everybody

12:32talk about it, but we have it here in my

12:35hands. It's a Nvidia H200. It's one of

12:38the most expensive part of this AI

12:41server. It cost about €40,000

12:44every graphic card like this. And all

12:47this rack cost more than €200,000.

12:51And why we invest so much money to this

12:54iron? The main reason is that it help us

12:58so much with the building competitive

13:00advantage in our company. The first

13:03thing is that we can use here customers

13:06datas and don't share our customers

13:08datas and many internal datas of the

13:11company with the third parties like

13:12entropic or open AI and I think that it

13:15build competitive advantage inside of

13:18our company. The second important part

13:20is that we collect datas on our own

13:23servers and in a long term it's more

13:26cost efficient cuz now we pay one of

13:30cost for hardware like this and we own

13:32this hardware and we don't pay for

13:34tokens. Many heavy users in our company

13:38use API tokens for thousands of euros

13:41per mount and I see high rentability

13:44especially on IT department where our

13:47colleagues uses so much tokens. I expect

13:50that in the future it will be market

13:52standard that big companies will have a

13:55local models and use it for internal

13:58datas and for heavy users uh in a

14:00company and I think that it will be not

14:02only about security but also about the

14:05cost and costsaving. Another advantage

14:07of this is the scalability that on this

14:11rack we can buy always some new hardware

14:13and when we will see that our colleagues

14:16use it so much after that we can buy

14:18more this uh Nvidia H 200 and put it uh

14:22here to this rack.

14:29The reasons why we have a separate uh

14:31room for the servers is of course

14:34temperature that here is much colder

14:36than in other rooms but uh also noise

14:39that this Nvidia H200 it's extremely

14:43noisy and it's impossible to work you

14:46know now it can hurt

14:53I think that uh if you are in the same

14:55room you can talk and work But it's not

14:59only on the same level that when we put

15:02large tasks on this uh server after that

15:05it's much more noisy if we have don't

15:07have so many requests it's a little bit

15:10calm and now I can show you how it work

15:13the first thing when we implement AI to

15:16our company was very basic solutions

15:19with a CH GPT and we use API from CGPT

15:23and after that we started using another

15:25model like a deep l for translations and

15:27now I think that we are in something

15:29like advanced mode of using it. Now we

15:33have a combination of online models,

15:35local models and for me like a

15:38stakeholder. It's very useful AI agents.

15:40Here it's our select. It's our main

15:42communication channel and here I can

15:44write some prompt to Bob AI. Bob AI it's

15:48our agent who responsible for reporting

15:51and I can send him a link and ask about

15:54reporting or send to him request and he

15:56can solve everything in a few seconds or

16:00minutes. Another interesting part of our

16:03text stack is that we have our own

16:05models and uh here we can see that in

16:07our own interface we have a four local

16:10models to are very general intelligence.

16:13The biggest is this open AI GPT. This is

16:16something like a general model for most

16:18of users from the company. And another

16:21model is Gwentry and this model use our

16:24IT department, our software engineers

16:26for many use cases like code reviews or

16:29they generate code from this model. And

16:32I think that future of a work with AI

16:35it's not only about this agents, not

16:38only about the prompts, but it's about

16:40the AI loops and AI loops are very

16:43simple solution. I think that many

16:45non-technical users can use it in their

16:48job and I can show you example one of

16:51them here I have a desktop version of a

16:53quot and you can open part coork and

16:56here I have a very simple AI loop where

16:59I connect my communications from my

17:02social medias from Gmail from my Slack

17:05communications and this AI loop every

17:08day collect datas from so many sources

17:10this AI loop analyze the most important

17:13things what will share with me and every

17:16day in the morning share with me the

17:18most important things and I develop this

17:20AI loop in a less than 5 minutes and I

17:24think that it's not necessary to be

17:26engineer for develop loop like this

17:28because it's very simple you can as a

17:31code how to do this and I think that it

17:33can help you so much cuz it's always do

17:36some jobs and this AI loops are not a

17:38new approach I remember that 15 years

17:41ago many companies run RPA is robotic

17:44process automations and this was usually

17:47a huge project in a company when five

17:50stakeholders need communicate with the

17:52five engineers about half year and after

17:55end of this job there was some bot who

17:58click something in a loop but now it's

18:01much more easier now everybody can

18:04download in a few minutes code a desktop

18:07version open

18:09and create AI loop without software

18:13engineers ers in a few minutes. I think

18:15that it's so easy and everybody can do

18:17this in a 3 minutes very easy AI and

18:20many experts think that this is future

18:22of AI that AI will work 24 hours, 7 days

18:26and week without human and many expert

18:29think that human is batonic of a system

18:32in a AI. What do you think is human in a

18:35batonic? Please write me to the comments

18:37and I will be very happy if you will

18:39share practical use cases of AI whoops

18:42or AI agents in a comment section. It

18:45can be very inspirational not only for

18:47me but for many watchers of these

18:49videos. Our implementation culture of AI

18:53in a one sentence is that it's much

18:55easier to ask for forgiveness than for

18:58permission. And I think that helped us

19:00so much that from day one I used to hire

19:03so many software engineers in a company

19:05on a non-technical department. We have

19:08so many technical people and it's

19:10automate our company bottom up. And what

19:13do you think what will be in the next

19:15few years automated by the AI? What can

19:18be automated by the AI and will be very

19:21necessary human touch in this field.

19:23Please write me in the comments and I

19:26will be very happy if you will share

19:27this video with the relevant people.

19:30Subscribe this channel and I'm looking

19:32forward to you in the next video. Bye.