EP. 1 AI for IT: Accelerating Network Deployment and Improving User Experience

0:07 we're going to go ahead and get started um thank you again for joining we are so excited to have you all in attendance today um before we kick things off I do

0:14 want to take some time to introduce our speakers uh today we have Satish Kumar who is a senior network manager at

0:21 service now as well as navraj panu who's the director of data science here at

0:27 Juniper um and my name is Selena gatis uh I am an AI Ops uh pmm here at Juniper

0:33 and I'll be the moderator for today's session um so jumping right in during this webinar we'll talk about an

0:40 impactful AI powered feature that was developed with the help of our friends at service now we'll show you what's

0:45 behind the AI curtain exposing how AI Works which is also known as explainable Ai and how Juniper developed this AI

0:53 capability to improve the experience of something most organizations use today and coincidentally what we're using

0:58 today to host our webinar um zoom and then finally we'll have some time at the end for a brief Q&A so please place them

1:06 into the Q&A section throughout the webinar um and then also during the

1:11 presentation we'll do some fun trivia questions just to test your AI history knowledge um so let's go ahead and get

1:18 right started with understanding Ai and networking uh Na Raj I'm gonna direct

1:23 this first question to you uh as a data scientist can you describe your dayto

1:28 day and can you please share what goes into building an AI sure um I guess the first thing that

1:36 comes to mind is we need a question we need a question to answer so for example in this

1:43 particular uh topic my question was can you predict if your Zoom call will be

1:50 good right so it's quite a general question then we have to see if we can

1:55 find tune it and have data in order to train the model so how did we fine-tune

2:00 that question we asked can we predict the latency of a zoom call or can we

2:07 predict how many times a network packet loss will happen during the zoom call so

2:12 that's something that's measurable and we can then Define a metric to Def to see how well our model is doing for

2:19 example can our model predict the latency of the zoom call right and the

2:25 nice thing is is that um Zoom provides us this inform information so we have

2:31 labels Zoom provides us with the latency for um the audio video and also if there

2:38 was packet losses so then we can develop our model and see how well it agrees so

2:44 we have our first metric how close are we are we within can we predict to

2:50 within for example 20 milliseconds uh the latency for 90% of the calls right and that that's actually

2:57 the metric that was uh agreed by the PMs and that's what we did and we succeeded

3:04 right so then the next question is as you asked how do you actually develop this model well you have to think about

3:11 all of the important things that go into the model and if we have the data the nice thing is we had the data from zoom

3:17 and we have our Juniper Miss data all in the cloud so we can combine them so what were some of the features or network

3:25 parameters that we use in our model well one thing is are we too is a client too

3:31 far away from the access point right that's a measure of the Wi-Fi signal if it's poor chances are or you might have

3:39 a poor experience are there too many clients on a particular access point is

3:44 it too crowded does that affected all of these are open questions no one knows right no one knows the answer these

3:50 questions and that's one of the exciting things about developing a model we can answer these things or you know for

3:56 example is it um a site problem is you know just there not enough bandwidth or latency at the site level that is again

4:04 another Network parameter that we use so we take into account all of these features or network parameters that we

4:11 have we have tons of data millions of data points for that and then we can develop a model so what model do we use

4:18 you know we've heard about chat GPT and you know some of the Specialists might have heard about Transformers or neural

4:24 networks or tree based methods you know those are all open to us and the nice thing about data science scientist

4:30 that's where the science comes in you experiment you see what is best you look at how much data that you have for

4:36 example if you have millions you can have a very complex model so some people

4:41 might have heard how long it takes to train chat GPT like a month uh on you know a dedicated cluster of gpus right

4:49 they have tons of data you know our data is a bit less but you know still it's a significant amount of data if you look

4:55 at the Miss Universe so we can develop very complex models so and that's in fact what we did and always the hard

5:02 part when you're developing a model the hard part is getting the data one of the hard Parts the second is is it actually

5:08 working and that's where you know predefining the metric for example how close we are to the

5:14 latency that is an indication of its performance and if it's performing well

5:19 that's great and we have a good indication about going forward with it

5:24 but we go back to the question can you predict if the

5:30 uh if my network will support an audio uh call and can you explain what could

5:37 be the problem and that was a big uh issue here can we actually explain what the problem is so for example was a

5:44 client too far away from the access point so that's the second part once we have a model and that we can trust it we

5:51 go to the explainability there's many different algorithms also to do this so

5:56 there's for example you can check out search lime uh and explainability AI

6:04 there's you know you can remove one of these Network parameters at one at a time and see the effect it has on the

6:10 model the one that we had chosen to do is called chapley it's a industry standard used um quite a bit in for

6:18 example the fraud industry trying to determine what are the parameters for fraud and um also in any many different

6:26 other types of predictions so with this shoply then we can see what is the

6:32 dominant reason for our call failing so was it a client problem it could have

6:38 been the particular um model for example device that a client was using that's that's a

6:46 possible reason maybe there was a client um configuration that was used that led

6:52 to this problem as long as it's a network parameter in our model we'd be able to predict it or for example if

6:58 there was just not enough latency at the site level to do that so when all of

7:03 that is done and validated and we go through quite a different uh quite an

7:08 extensive validation process at Juniper where the project managers come in and luckily we also had Satish come in and

7:16 say you know this H this makes sense so which is uh anytime we can get dedicated

7:22 customer uh critical feedback it's it's gold so and then uh once we pass that

7:29 we push it out to production and also equally good I work with great colleagues who can scale up this um the

7:38 models that we've developed make it generally available it's awesome sounds like tons

7:44 and tons of work that goes into it none enough words in the world can explain it huh

7:49 um so let's before we jump right on into our next session here uh let's run a

7:54 quick poll or trivia question if you will um I will go ahead head and put

8:00 that up the one was the concept of artificial intelligence first

8:06 coed I'll give a few seconds here see some answers coming

8:19 in hey

8:24 so all right so if you answered we can go over to the next

8:32 slide awesome if you answered 1956 you are correct so gold star for

8:38 you um great job let's move on over to our next

8:44 section here uh so fatish let's let's talk about the AI

8:51 solution benefits for service now um so s why did you choose Juniper and what

8:57 are some benefits you've achieved using our AI solution absolutely um let me start

9:05 about maybe about five years ago and service now we started with M Wireless

9:11 then it was working fine great right we saw a lot of the wireless issues what we

9:17 had before Miss migration then about two years ago uh

9:22 when the jiper acquired Mist right that's where we we we started thinking of um uh moving toward the switching and

9:30 also the Juniper SSR which is a 128 company which is acquired by Juniper

9:36 right so two years ago we started with this migrating to switching then SSR so

9:42 as a last year we are like a full full stack Juniper uh Wireless wired and van

9:49 so why did we choose Juniper right so we have mainly I'll put it in the three

9:54 ways right so the four four reasons mainly one is simplify the network

9:59 operations and second one is bring the end to end visibility

10:05 and the third one is bring the better user experience and the last one is

10:10 obviously the cost Factor right so if you look at the first one the simplified network operations what I mean by that

10:18 uh before Juniper it was very challenging uh for a network team to do the operations right either the code

10:24 upgrades or the board configuration or vand changes right all the day to operations it was never easy and

10:32 automation capabilities is also very limited so that is where the Juniper platform helped us to simplify right how

10:39 did it simplify so it is easy to um provision right using a ztp zero touch

10:46 provisioning is like a very good one right it's we never had before that like that I always call it as a true ztp it's

10:54 a cloud-based just like a scan barcode and all it need is a d CP IP address

10:59 connectivity and rest everything is pushed from the Mist right so it simplified overall provisioning process

11:07 even bringing up new sites it was never easy before so now we don't do any travels much so we pretty much do all

11:14 the uh the work done remotely by the network Engineers having maybe local

11:19 contractor do the raken stack physical work things like that right so that's on the uh Day Zero part

11:27 and also there's another one important U which is the templates which is I would say this is like a great great

11:34 capability we always had a challenge on maintaining the network configuration

11:39 standard right it's easy I mean any network engineer they go through that we

11:45 Define the standard in a spreadsheet but maintaining the spreadsheet is never easy and it'll get lost over the period

11:51 of time so what we Define in the spreadsheet and uh it's never be the same on the um on the actual Network

11:59 devices so this template is really simplified that and we also have the

12:04 compliance compliance we need to make sure we follow certain guidelines on the configuration right so the templates it

12:12 greatly not only simplified and it also helped us to maintain the network

12:17 configuration standards throughout the life cycle so now I can tell confidently

12:23 without even logging to any device by by going to the templates I can tell hey what is our standard and U the Mist

12:31 takes take care of all the magic right we don't have to log into devices we don't have to do any scanning everything

12:37 everything is techn about it so that is another second good feature and the third one the next one is a dtp dynamic

12:45 uh Port profiling it so this is also a great feature because most of most of

12:52 the operations work involved configuring the ports things like that I mean the new devices comes and goes away right we

12:58 got always the help team comes to say can you configure I I got a new printer

13:04 or I got a new security camera or I connected in this so support can you configure the vand so all that work is

13:11 been taken taken off now by the dynamic Port profiling how we can do the

13:17 profiling in the Mist portal so then it becomes like a more of the plugin playing right you once you do the

13:23 profiling for the devices so they can connect the help team or the support teams they can go back and they can go

13:30 to the IDF and connect the devices and this automatically configur the port without even uh network engineer to do

13:37 anything right it it works just like a magic so some of these features help to

13:42 simplify uh in the network operations right this is by by default by the mist

13:48 and at service now even we took a little bit further a step right even

13:54 um for example the uh the firware upates things like that it was super easy like

14:00 two clicks select the firmware upgrade the version then second click upgrade

14:05 and this everything is taken care by the uh the cloud right M Cloud so service

14:10 now even we have taken for the next level we have workflow this entire

14:15 process end to end uh automation workflow for the firware upgrades that

14:22 includes the validation part so now we don't do any manual upgrades we we don't even do

14:30 no including the validation is like a full end to end automation workplace it's been working great how this is all

14:37 possible because of the good Rich apis by the supported by the mist and having

14:43 the integration between two platforms to Great platforms we're able to do some of this great

14:49 stuff so another other important f is the visibility right that was another

14:54 also a very key so we always have a problem in u when things break troubleshooting is

15:02 always challenging it was never easy so we wanted we always uh wanted have good

15:07 visibility on the data that was a missing piece so when we decided to go

15:13 with the full stack Juniper so that helped to solve the problem so now we are since we are full stack Wireless

15:19 wired and Van we have full visibility in the entire network the packet goes from

15:24 the wireless wire and the van Network and having this capability of junifer

15:31 the marus missed so it can on top of that on top of the network you have the

15:36 a capabilities which is a marage so it is not only telling up down the switch

15:42 down things like that it does beyond to right it kind of predicts the network it it gives a deep visibility and it gives

15:50 the the Marv actions whether it's a DNS issues or DHCP issues or the

15:55 misconfiguration so some of those all the stuff is is out of the box right we don't have to even look into that so

16:02 overall troubleshooting has become very easy and bringing this the end to end

16:08 the visibility the data is what helping a lot and combining all these features

16:14 is what it is driving for the better user experience we want to give the

16:19 users the best um Network experience right even when there's a problem so

16:25 users need to know immediately right we're trying to move away from a reactive approach more of like a

16:33 proactive right we're trying to give some of the enable some of the features like a self service users themselves

16:40 they can troubleshoot the issues by themselves by having this having the integration between service now and the

16:48 jiper missed so we we can L some of those capabilities and let the users

16:53 troubleshoot the issues by themselves before even they reach out to their help do things like that

16:59 so yeah combining all this that's how um these are some of the reasons why we picked Juniper and uh this is where we

17:07 are this journey to that is awesome so happy to hear that we are we've checked

17:12 all the boxes and we continue checking them um naaj I don't know if you wanted to say anything before I jump right on

17:18 into our next trivia question or we can wait until next H yeah we can wait until

17:24 the next okay um so our next trivia question

17:30 here let me go ahead and put that on the screen uh what is the primary objective

17:35 of reinforcement learning in AI I'll give it a few seconds

17:57 here all righty looks like people are changing their minds putting one thing

18:03 so I'm gonna go ahead and end it here's some results if you want to

18:09 move over to the next slide if you answered learning from trial and error

18:14 you are correct so another gold star for

18:20 you um let's keep going so we're going to move over to this next section which is

18:27 sort of like the meat and potato of this webinar um be a two-part question I'll

18:33 ask something to Sati and then nav feel free to like chime in wherever um so Satish uh tell us some challenges you've

18:40 had with your Zoom calls and then navage you know how did you develop this integration to resolve those challenges

18:47 so Sati I'll go ahead and give you the floor y so service now as Zoom is always the

18:55 the most critical application the service that heavily been used right if Zoom is not

19:01 working it's it's kind of really not a great thing right so maintaining um the

19:08 zoom application all the time uh healthy and make giving the best user experience out the zoom calls is super critical for

19:16 us so even with the juper when we acquired the full stack juper right we

19:21 always had this challenges with zoom complaints so users always complain that

19:27 um they been having issues with the slowness audio video choppy all all that

19:32 kind of issues right but when we troubleshoot these kind of issues it's

19:38 very hard right even with what whatever the data we have um it it tells okay

19:44 it's not a network problem but it doesn't tell enough that uh why it is a

19:50 problem with a zoom right so that that was a missing piece so our focus is not

19:56 just passing the ball even we trying to to okay even if it is not Network issue we wanted to tell where the potential

20:02 problem area could be right so that is where then we started discussions with the Juniper Miss team and we told them

20:09 okay let's do the integration with the zoom and bring the zoom data into M

20:16 Cloud right M portal so when we do this when when that integration happened so

20:22 now we can see the zoom statistics of all the users of all the calls in the same M portal so what does it mean so

20:30 now even when somebody reports an issue it's kind of very easy um to identify

20:35 the issue and correlate whether it is because of the network issue Wireless

20:41 wide or V or is it more on the application side with is which is a

20:46 zoom so and it helps with the the correlation too and most of the time

20:52 it's what based on our experience 80 90% of the time the issues relies on the

20:58 client side right the laptop side laptops maybe High CPU or high memory

21:03 the resources running very high on the laptop those are the contributing

21:09 factors for the bad uh Zoom call experience so with with having that Zoom

21:16 integration with the McLoud bringing that data into the same one place with

21:22 the correlation helps uh so much easy for troubleshooting to give an example

21:29 we had a couple of months ago in one of the site in us um it was in Orlando right so users used

21:36 to complaint the zoom quality issues but after done the thorough troubleshooting

21:42 we found out it was due to some hardware issues and also the Wi-Fi interference

21:47 and height of the aps so then um we identified and we fix

21:53 the problem then how do we measure whether did we really fix the issues or not not right after the upgrade after

22:00 adjusting the height of the access point all the stuff so issues got settled down

22:06 right so issues automatically significantly drop right no not much

22:11 complaints on that then we wanted to see by ourself hey is it really improved or

22:17 not right that's where even the the Jer Miss team help using the AI ml algorithm

22:24 the back end they were able to provide some of the stats that tells how the experience was before the changes and

22:32 how the experience was after the changes so we are able to uh um see almost a 40%

22:39 Improvement in the zoom call overall experience based on the data that's been

22:44 provided by the M IML

22:49 team yeah and this was a a great uh example where

22:57 you know the customer and Juniper Miss got together service now Juniper Miss got together

23:03 and tried to solve this problem like and again you highlighted some very important things the whole idea was

23:09 trying to identify the problem right you know is it a client problem is it a

23:15 problem with the access points or is it a site problem you know these the these are all very sensitive issues and you

23:22 know everyone wants sometimes maybe wants to point the finger at everyone someone else but you know and that's

23:28 nice thing about the transparency about the explainable AI right so I mentioned the shle values beforehand so these you

23:36 know this is not something that we uh manually code in and say oh you know it's always a client problem or

23:42 something like that it it shows us uh based on you know it's mathematically proven even given a Nobel Prize for uh

23:50 the fact that it um uh gives a fair unbiased and objective

23:56 um contribution to whichever one of those factors are so and uh the the thing

24:03 about networks as well it's very Dynamic so things can change and they do change for example after your after the

24:10 successful upgrade right we we see the performance going on but the nice thing is we have this measure you know coming

24:17 from Zoom about the latency or the packet loss and then we can measure to

24:23 see oh did something happen next week uh or something happened last weekend now

24:28 we have a huge performance uh decrease so uh just just as another example um we

24:36 had noticed that um Zoom performance was significantly uh poor all across service

24:42 now so I I approached uh the PM um Kumar

24:48 putas Swami and I said you know what's happening so I think Kumar talked into you sa and he said oh there was an email

24:56 that um people are going back to work so and we saw that immediately so you can

25:02 see it and then you know that and just as Sati was pointing out you know

25:07 we we can proactively uh solve these problems and right then we know what to

25:13 expect and we can address them so and then having this data gives it all this power and then also with the dashboard

25:19 you know uh Sati knows what's happening as well and then getting that information to the customer showing the

25:26 troubleshooting and seeing that it's subjective because you know uh I can only imagine you know changing access

25:33 points or lowering the roof uh you know it's it's a very costly Endeavor you

25:38 know and you want to have it reliable right like what if we were wrong what if it was something else right then that

25:44 makes uh everyone look bad right uh but you know that's if you rely and you know

25:51 have proven metrics that um have been quickly valid or correctly validated

25:56 then we can avoid these issues so it's a great story yeah just just to add on

26:02 that n what you said is absolutely right because we need to make sure after we do the corrections after we fix the problem

26:08 right we need to see the results and we know that there are maybe another one or two sides we have something like that

26:15 but I was not confident hey I mean replacing AP is changing the the

26:20 building the structure the cabling it's not easy it's painful right and that's we have to depend on the remote remote

26:27 um site contact it was never easy and I was not confident until I see the data

26:33 from you okay after we fix all the issues now I was confident that okay we saw the Improvement in the zoom quality

26:40 overall experience right so that gave me confidence even going forward even if we

26:45 need to take this path for other sides we can do it now awesome um great to hear that the

26:53 clear communication was a big one and the success too um before I move over to the next section I

27:00 did want to put another disclaimer out there you know if you have any questions please put them the Q&A section here on

27:06 your screen um I believe we have one more trivia

27:12 question let me bring this up

27:19 and should be on your screen here so which breakthrough AI technology

27:25 achieved human level performance on a broad range of natural language understanding task in

27:31 2020 if you do not know this you might be under a rock

27:39 um I'll give a little

27:51 bit okay right it looks like majority answered at this point I'll go ahead and

27:58 share the results if you want to move over to the next slide if you answered

28:05 gpt3 you are correct um so if you've got all three of these right congrats three

28:11 gold stars for you um and we do appreciate your your

28:16 engagement so that being said let's keep

28:22 going so fatish now that we're towards kind of like the ending of the webinar

28:27 here um being a valued Juniper M AI customer one who have helped us um improve our AI

28:33 solution over the years uh what is the capability you like to see

28:39 next sure um since the M Marvis has air

28:45 engine it knows what's happening right it has a full data the complete visibility into the network and it also

28:52 kind of predicts some of the stuff right so I'm looking forward for the more some of of the the selfhealing capabilities

28:59 in the network level right it does maybe some little bit on the wireless probably

29:04 it's good to if we can expand further into the wireless I mean uh further more

29:09 from the wireless right switching and also the um the van Network and also

29:15 giving the proactive notification I'll give some of the scenarios for example if there is the issue with one of the

29:22 van circuit right not hard down but maybe the Laten issues or uh um maybe

29:28 the Jitter some of this the bandwidth concerns things like that so maybe this uh self healing um capabilities should

29:36 automatically fail over to another circuit right without doing any manual

29:42 intervention and also some of the proactiveness so since we have all the

29:47 data maybe it should predict some of the issues what's coming for example if the

29:53 user is have a scheduled meeting in one of the flo floor and um if Marvis says

30:00 there is some degradation in the network in that floor maybe should notify the US

30:05 user proactively saying that hey there is some uh Network issue kind of

30:11 probably and we we know that you have a schedule meeting coming up in this maybe a a side of the building maybe it's

30:19 better off you move other side of the building to different conference room so that way this kind of bringing the

30:25 proactiveness and the self filling identify the issues and proactively fix

30:30 that without even network engineer or operators being involved yeah so the

30:36 some of the things I'll be looking forward to those are all excellent ideas uh for

30:44 like I mentioned we have all the data now and you know I would actually just like to applaud service now and Satish

30:50 because they provided the data you know they trusted us with their data their data is secure in the cloud they did it

30:58 before Juniper did it right so we are able to uh have the data and play with

31:03 this before and it's just a real tribute they they you know they they you Sati

31:09 told us you know there's a problem here where we're having poor Zoom performance right and Juniper listened and uh the

31:17 Juniper Miss team basically hired me to solve uh this problem so and uh you know

31:24 just making the data available because you know we are not the the first step

31:29 you know just to Selena's first question you we we had the data because we had

31:35 the data we could do it and then uh we could solve these problems so and then you know all those things you mentioned

31:42 uh I was just noticing we have data points that can tell us that right so we can add them to the model so having this

31:48 feedback is um you know essential so please keep it going because you know we

31:54 test it and then we see does it actually solve the problem so um I'm looking forward to uh version two of the model

32:03 to to account for these things as

32:08 well um well thank you Sati for sharing that with us again we are working

32:13 towards that um so now that we're kind of towards the end of the webinar looks

32:19 like we have a few questions here um if you do have any questions again please

32:25 pop them into the Q&A um section there the first question we have is how do you

32:32 determine the efficacy of AI um and Satia I'm gonna actually

32:38 direct that one to you yeah

32:43 um the efficacy of AI right so we we heavily um depend on Marvis right even

32:50 at service now we have our own AI model but we're trying to not reinvent the whe

32:57 so use use the data what Marvis says and um on top of that we try to build our

33:03 own AI model right so what that mean so Maris um some of the times it provides

33:12 OT very good information right very good insights and we take the data right even we convert some of the data into a

33:19 service note ticket so that way we don't just see the data we take an action to

33:24 that so with some of the Integrations the direct integration with the Juniper mist and the service now so we

33:31 have that capability whatever Marv actions say that so we convert into service no tiate so we look into that

33:38 action hey is Marv's action telling right or do we see any discrepancy right

33:44 and uh because we validate I mean it it's good it is right most of the time but sometimes it is not getting the uh

33:51 right information that is where we provide the feedback to the Juniper team hey Marv is is telling maybe this bad

33:58 cable maybe uh two out of 10 cables maybe not the real bad cables right so

34:03 we give this data back to the m team so they run uh this data they evaluate and

34:10 come back with the feedback and try to solve this so this is kind of more back and forth like share we see the data

34:16 review the review the data and we share the feedback to J per Miss so they help

34:22 us to solve this I think it's kind of a journey I don't there's no clearcut um

34:27 step and it's a model right we have to train we have to train so we have the

34:32 data we have we got to we got to continuously um train it and improve improve the

34:40 product um we have another one here that says how long does it take to develop an

34:47 AI feature um I'm gonna go ahead throw that one at you nav you it's a really

34:53 good question so um I mean the the hard part is actually do we have the data and

35:01 then getting the data um it usually takes uh it can take sometimes if we

35:06 have the data for example um you know the schedule for a zoom call and if it's

35:11 uh occurs all the time um I imagine that data is there so we could have it and

35:17 then we could train the model so but then you know uh we add the feature in

35:22 to the model and then we test it has it improved it does it show up so and that's actually um going back to the

35:29 first question that was asked the efficacy right is it able to point it out or is it able to stand out where it

35:36 should right so the these are the sorts of things that we test so first getting the data and actually seeing if it's a

35:42 useful uh feature and does it show up when we think it will show up so um the

35:48 the whole process actually can take you know maybe about two to three weeks to do the whole uh data Gathering and uh

35:58 validation aome um and then I think we have room for one more uh so we have another that

36:06 says what's an example of an AI feature of the Miss solution that you and your team use daily and how has this improved

36:14 your operations so te I'm G give this one to

36:20 you um sure yeah so let me let me think about so in Marvis what do we use um

36:28 more frequently right Marvis actions are definitely a good one so we um Maris

36:35 actions and mql even I personally like the mql so I with the mql Maris sarey

36:42 language uh we can get the data very good actually I mean it has a data then if you need to generate some of the

36:48 report and it's very very uh useful and SL is the most important I I

36:55 look at that every day morning how is how is our Network looks like right

37:00 before my day starts just a few click the dashboard go there sles AR level

37:06 sles it tells you like how overall Wireless wide van the health status

37:12 across all the sides globally right that one and um Marvis actions is another

37:19 important one like as I was explaining before we heavily uh rely on the Marvis

37:27 because Marvis actions if if Marvis is telling something wrong with the DHCP we we really wanted to know what the

37:33 problem is right is it because of the connectivity issues or is it because of the actual DHCP server itself right or

37:40 it can be a DNS issue so some of these things uh we uh heavily rely on this and

37:45 as as I was briefing earlier we have the Integrations done with service now so

37:52 every Marvis action we take it serious and we try to review the data and if

37:58 there is any discrepancies we try to correct and by working with the Juniper

38:03 Miss team Marvis actions mql slus are the

38:09 most uh daily used features though thank you for sharing um I

38:17 believe that is it we have time for for Q&A I do want to share some resources um

38:24 with those who are interested uh so if you want to either whip out your phone now and just scan away or you can take a

38:31 screenshot I encourage that as well um and that way it's already saved to your desktop and you can bring it up at another time um but I encourage you to

38:38 read more on um the Gartner report we do have a few demos here and a few um or a

38:44 web page here of explainable in AI um and yeah we would love to have you on a

38:50 future demo if you are interested and that is all for now I do want to say

38:55 thank you to SA for joining us as well as navraj um we

39:01 do appreciate you all's uh engagement conversation wisdom all that good stuff

39:06 and again thank you again to the participants for your engagement in the trivia and hopefully you feel good about those three stars um and yeah that is

39:14 everything we thank you and hope to see you on the next one bye for

39:25 now