AI-Native Telco: Genie and Autonomous Networks with Databricks

Summary

Databricks Genie enables telecommunications operators to break down silos between business and network teams by providing domain-specific conversational analytics spaces for customer care, billing, and network operations.
A supervisor Genie architecture orchestrates multiple specialized domain spaces, enabling root cause analysis and proactive network diagnostics that surface answers before users even formulate the question.
Databricks integrates with TM Forum's catalyst program to deliver standardized AI innovation, including live demonstrations of autonomous network Level 4 remediation powered by AI agents, MCP tools, and MLflow traceability.

AI-Native Telco: Genie and Autonomous Networks with Databricks

In telecommunications, network reliability and customer experience are sacred, yet business and data teams have long operated in silos. this video showcases how telecommunications operators are breaking down those barriers with AI-powered conversational business intelligence using Databricks Genie.
Learn how domain-specific Genie spaces empower business users to get answers in natural language, from real-time CFO dashboards to autonomous network management. Discover the architecture of supervisor Genies coordinating specialized domain spaces for customer care, billing, and network operations, how Genie provides root cause analysis and proactive diagnosis before users even ask questions, and how Databricks integrates with TM Forum's catalyst program to drive standardized industry innovation. See live demos of order-to-cash workflows, what-if modeling, and autonomous network-level-four remediation powered by AI agents, MCP tools, and comprehensive MLFlow traceability.
🤝

Chapters

FAQs

What is the TM Forum and how does it work with Databricks?

The TM Forum is a 35-year-old nonprofit organization and the largest connectivity industry alliance, with more than 800 members including Databricks, which also sits on the TM Forum board. The forum's catalyst program is a unique proof-of-concept initiative that brings members together to build and promote open, standardized industry innovations in telecommunications.

How does Databricks Genie support telecommunications operations?

Databricks Genie empowers telecom business users to query data in natural language across domains including customer care, billing, and network operations through domain-specific Genie spaces. A supervisor Genie coordinates these specialized spaces to route queries correctly, enabling real-time CFO dashboards, order-to-cash workflow visibility, and proactive network diagnostics.

What is Level 4 autonomous network management in this context?

This video demonstrates a Level 4 autonomous network remediation capability where AI agents detect anomalies, classify issues, and take corrective action on network infrastructure without requiring human intervention for each step. The system uses AI agents, MCP tools, and MLflow traceability to ensure the autonomy is auditable and governed.

How does the Genie architecture enable what-if modeling for telecom?

The supervisor Genie architecture connects domain-specific spaces to enable what-if modeling, allowing business users to ask hypothetical questions about network performance or revenue impact and receive data-grounded answers. Proactive diagnosis capabilities surface potential issues before users ask, combining real-time data with predictive analysis through the Databricks Data and AI platform.

Full transcript

[00:09] Welcome everyone. My name is Andy Flint. I lead our taco business for the US. I'm here with Guy Lupo from Forum. Hello. I'm Guy Lupo from the TM forum. I uh lead our trustworthy AI and data mission for the TM forum. Thanks, Guy. So what we're going to do
[00:24] is first talk a little bit about TM forum for telco and some of the outputs and then we'll follow that up with a couple of demonstrations one on genie and another one on model as a service. So um guy TM forum right obviously an important foundation for telos and you
[00:41] know across the globe. Can you just talk a little bit about um what what it's about, right? And some of the participation and and things that uh that telos benefit from. Easy. So, for those of you who don't know the TM forum, um and whoever's been
[00:59] in Telco around knows it. Um it's been around for 35 years. It's a nonfor-profit organization that is probably the largest connectivity industry alliance. We have more than 800 members amongst which data bicks is one. Data bricks is much more there also
[01:15] sitting on our board. I will mention that in a second as well. But um um with more than 135,000 practitioners around the world, multiple conferences, collaboration activities and what we do is we have this very large open digital
[01:30] architecture framework that most of the telos have been using for a long time. names like ETOM, seed and other frameworks that are being used to futureproof, design and integrate uh save cost, reduce opex and uh that's
[01:46] what the the forum is I guess. Thank you guy. So uh data bricks is a pretty big contributor in TM forum. Uh we're on the board we do quite a bit. Can you tell us a little bit about the catalyst programs? um and you know
[02:02] what are the what are the outputs of that and how do customers access those uh assets. Perfect. Yep. So um TM form catalyst before I jump into catalyst just a bit of um um um where it all sits um we have many activities we collaborate
[02:19] uh we have innovation labs uh and we have research and also we have a proof of concept um program this is very unique proof of concept program is called the catalyst program the catalyst program is uh
[02:36] special it's under the bylaws And what it means basically if you come into uh TM forum um like data bricks came um and say I have got something special which I would like to actually create for the industry with the
[02:51] industry to promote openness to promote the standard and to create an opportunity to push you know push the needle forward. We will then go and find more willing parties and then we assign a success manager who goes and brings in
[03:08] 10 other telecommunication providers you've never met before and together they come together and put a proof of concept program which goes all the way down to the coding and then that basically gets celebrated not to mention gets to appear as a booth
[03:25] in our conferences wins awards and made available for all the members and non-members if they register to follow the project um at some aspect to actually get to use those resources the architecture the reference implementation and if someone is really
[03:44] uh precious about it we will even allow you you guys to come into our innovation hub and join that program to actually experiment with this in code and build upon this forward. Great. And uh how do you see that moving forward right as we go into the future
[04:00] to continue to deliver value and um you know across across the globe with the telos? So to the catalyst program is been running for a long time. I think there'll be about 75 odd catalyst in next week in Copenhagen in our event which is our flagship event. Um
[04:18] to make it move forward it's not enough to do pox. the the trustworthy AI and data mission which was born 18 months ago and is the third mission out of to others composable IT and autonomous network and trustworthy is me. We we
[04:33] thought that we need to actually provide the members in this day and age of vibe coding and using uh coding agents and software development. We have to vote with code. We stole that sentence from AWS by the way. We love it. Vote with code.
[04:49] Okay, which means we are going to do stuff in code and to do that we came to our catalyst members and these catalyst members we told them you we're going to take it to the next level. So we've taken uh our canvas environment. TM form has a Kubernetes environment which has
[05:07] many components and open APIs more than 2 million downloads and users around the globe that allows you as a telco to talk to your billing system without knowing which billing system it is to use inventory without worrying about it. We
[05:22] realize that people want to go and build agents for telco to integrate easy to customer to do customer churn fault management and so forth. So we've called upon the industry to come and build more operators for our environment and that's where data bricks joined in and brought
[05:37] in data bricks and built a data operator for the industry under TM form canvas to allow to start working with data as an agent on our platform and that's a very big contribution and a very significant
[05:52] one. Okay, thank you for that. I think what we'll do now is show a couple examples of outputs from the catalyst program. Yep. All right. Um so Justin All right. That works.
[06:08] Thank you guys. So TM form and data brick. So I'm Justin Michaels. Um I run our field engineering organization in the communications industry. Uh uh Andy is my counterpart and we'll get to Scott Davenport and
[06:24] Leila Yang are both on the team. Um they're um they're solution very senior solution architects uh that are part of the team and um so they've done quite a bit of work as far as uh trying to show some demos that we'll show in a second.
[06:39] But real quick, like Guy mentioned, there's 75 different projects that'll be shown next week. We're invested in effectively three of them. Um so project foundation is the what the term the term project foundation is just the foundation for where a lot of these
[06:54] projects are run and built on. Um so we're a heavy investor in that project as far as where do we actually go and you know build a lot of these projects. Two of those examples are insight x and uh model as a service. So what we did is we've taken each of these
[07:11] um so first insight x is a genie based uh example for uh talking to your data. So that's one of the things that we'll be showing in next week in uh uh copen. Uh the other is a model as a service. So how do I secure and provide models out
[07:28] to my uh out to my you know different business units in effective way um without just having sprawl and everyone just creating their own models. Um so two examples are things that we just we want to go through. Uh we the next thing
[07:43] that we'll see from Leila is a example of a a Genie demo very close to what we'll be showing next week of Insight X. And then for model as a service we actually applied that to some of the demos that we have for uh an autonomous network example. Um and it touches on
[07:59] some of the model as a service u ideas that we'll be uh presenting next week. So we thought we'd go through those today and I appreciate your time. So Laya, if you want to take over for me. Thank you, Justin. Hello everybody.
[08:16] So Genie, if you went to the keynote, we are all over Genie this year. Well, Genie is definitely a keyword. Um, what I'm going to show you today is well, first it's the architecture here. Um, telecom has a lots of data. You have
[08:33] data from the order, data from the network, your customer data and telecom also have people like all of yourself that want answer from the data but all of us I'm part of it. We are not getting the answer fast enough and Genie changes
[08:50] that. Genie is the AI that knows your business. It sits on top your open format your open lake and governed by the Unity catalog. It connects through the Genieontology which is the key um announcement that we did yesterday. That
[09:07] means when Genie sees your data in the lake, it's not only just the data content itself, not just the metadata, it also sees the sentiment, the meaning behind your data and also the relationship that can provided by the
[09:23] ontology. So as a result when genie answer your questions you ask a genie about your business it's grounded your answer is grounded on your data context and grounded on your business um logic
[09:38] so that's really powerful imagine you're the VP of operation want to answer want to understand hey why my order dropped 20% yesterday and today um we file tickets to your analytics team an
[09:54] analyst to write a query and when it comes back um it's probably within a day maybe a week so the time has passed so Genie changed that scenario and this oh sorry can we
[10:10] okay nothing happened um this is when it's gets really powerful uh for telecom you don't just build one genie um for the entire company you build multiple domain specific genie that understand um
[10:28] your different domains so they they can embed it in your data um your data sentiment. Customer care genie for instance knows your contact center data. You asked the genie in plain English um
[10:43] about your customers inquiry why my bill went up 20% last uh last last month. network genie knows everything about your knock center um data so that it can triage the question when there is a error uh the network error error outage
[11:00] it knows how to triage this and my colleague uh Scott will show you a deep dive demo that genie and multi- aent um charge the network era using plain English that's really powerful and for for instance the building genie sits on
[11:17] your billing system and understands and say your your rate patterns and your distribute uh your dispute and your revenue recognition and so on. And what's even more powerful is on top of all the domain genies, you have a super
[11:33] you can put a supervisor agent to coordinate with different small domain genies. So when your CFO asks a question that spans a different business unit, you can uh the supervisor agent can consult different domain genies um
[11:50] digest the response and you know um and then cook one co co uh co cohesive response to you. So that's why the uh supervisor genie can serve as like um crossbu brain and understand different
[12:07] domain context. So this is the architecture. I'm not going to bore you with more slides. Let's see. Let me show you what it actually looks like. Uh when a c when a telecom executive sits down,
[12:22] start asking questions. Imagine you runs a telecom that has a more than 1 million a 100 million subscribers. um just thousands of orders and hitting your system every hour and you have
[12:38] phone activating you have fiber fiber install scheduling and um it's very difficult to consolidate all those data and if you are CFO this is let's us start with the group performance what we are seeing is here is a live platform
[12:56] every tab every metrics and every AI response is grounded to your business data. Okay. CFO um sits down every morning. This is your homepage, your morning briefing. Um without any without
[13:12] any analyst, the Genie agent can give you real time in the last 24 hours, your real time snapshot of your business. Um you know, 125 billion um in revenue.
[13:28] Well, 2.7 yearover-year growth. It's really good cash flow 40 billion and your fiber growth at a double digit. So the business looks really good. But hey, look at this. Your genie agent actually
[13:45] flex something for you while you're sleeping without anybody asking this question. the business business while line margin pressure it the operating um loss widened to negative86 million and further it it just do the
[14:03] diagnosis right it's driven by Lexi services revenue decline of7% um this alone is powerful because this genie is doing the diagnosis while you are sleeping nobody asked them to do so
[14:19] but even more more powerful is um genie one layer deeper it has a genie research agent it can cross reference to another domain um domain context postp turn rose to 1.05% 05% as device financing period
[14:38] expire and the suggests action is to accelerate converted the wline fiber bundling to reduce turn and shift the B2B account from legacy to fiber before the copper shut off deadlines. So this two genie um is just one example that um
[14:57] genie combine different domain genie combine can do this you know tracing and reasoning for you before you even even um start asking the question. Okay, when the CFO sees this during the morning coffee before morning coffee even finish
[15:14] um maybe the question becomes hey where my money um it actually it's stuck actually so let's follow an order through the system see the order to cash tap um I think
[15:30] what what every single dollar of the 100 125 billion start as an order and actually most of people don't even think about is every day. That's telecom revenue is decoupled from the order volume. Um because the average quote to
[15:48] cash is about 30 days to to 90 days. The revenue won't get realized um until your service activates. Right? So look at the scorecard here. Provisioning is really good. Building is fine. It's on target.
[16:04] equipment installation and digital order percentage needs to catch out a bit. But what's bit concerning here is to order to activate the activation is 3.2 days. Um well it's it should be two. So it's
[16:20] too slow. Um let's ask genie why it's too slow. What's the main reason for the delay in order to cache? So what it does here is the genie goes back trace back to the data pipeline and look through your table and look through all the data
[16:37] that's actually empower this this live dashboard and then we'll come back with a cohesive answer. This is a live demo so give genie a little bit of a time. Okay, let's see. B2B provisioning is the
[16:53] single biggest drag. Order takes more than eight days to oh even 22 days to activate versus 3.2 days um on average. So that's the the headline the the reason but the root cause is really the
[17:09] manual um the manual provision process and further you can see here maybe the font is a little bit small September 20 device launch expose the scaling failure. Oh Genie actually is aware of
[17:25] those anomaly events on the left hand side. New device launched September 20 98k orders. So the order warning actually search 69%. But on paper it it's good, right? I have
[17:40] more order hitting my system. However, it the provisioning queue dep. That just you know um that means your Lexus assistant cannot scale. Um and also the genie can is aware of the
[17:58] November 15 billing cut over which is this one. It the order queue oh the legacy system just cannot take the order. It has a cut off and order drop to uh drop 40%. So those compound effect
[18:13] drag down your activation time. This is the power of a genie. I want to call out take a pause here at this right moment. It just unpacked from this one number 3.2 days to the root cause analysis. So
[18:29] this is a transitioning from staring at your dashboard to you now you have a trusted thinking partner or the strategist to help you reason over your over your data. Okay. Um noted there is a disconnection side by side uh with the
[18:47] confirmed order volume because a CFO just want to see the in andout ratio want to see oh uh the other side of the the equation so that uh he and she can have a holistic picture of the order. So this is the uh order to cache
[19:03] and every single interrupt an interruption here will has this ripple effect to impact your P&L impact your business function. So let's take a look at business function. Overall it looks good. Um mobility
[19:21] 89 billion in revenue and operating margin is 30% positive. So this is definitely your cash engine business while line and the revenue is good but the margin is negative. Well know that
[19:37] we already flag this business wine business at the morning briefing the homepage. So this is the just like a more deep dive of the view. Look at the heat map here. Business wine confirmed in the last year the margin dropped from
[19:52] consistently dropped from negative 3.8% 8% to negative 5%. And the Genie research agent can pinpoint, hey, the key driver was fiber, but the fiber isn't grown fast enough as the legacy
[20:08] copper is bleeding out. But what's interesting right below here is the consumer wild line on the other hand, fiber is the key driver and being able to lift the margin from 8% to double digit to 11%. This is a pretty pretty
[20:24] impressive. So let's take a look at the consumer business how it looks like. Maybe we can learn something from there. Um the business looks good. The fiber is grow at a double digit. Well postpay and fiber appo both are impressive but hey
[20:42] this is something a little bit cons concerning here. Postpay churn increased 13% year-over-year. This is kind of counterintuitive because consumer business is doing well. So let's ask Genie why why postp pay turn
[21:02] increase. And again Genie is doing its reasoning looking at your customer data your order data and maybe even uh correlate with your network data to come up with answer. Let's see. Post Patreon rose 13%
[21:22] um because no proactive retention at installment plan expiry and further let's see the recommend action is deploy automated retention trigger six days before
[21:38] installment plan expire meaning um Mr. and Mrs. CFO, we definitely need a more effective um retention offer as soon as possible. Otherwise, our competitor just you know take away our um very valuable
[21:55] customers. So that's Genie answering the post term questions. But the real risk is well if your decline business like a legacy broadband your leg revenue outpace your growth engine your growth
[22:12] upside which is fiber. So next question becomes how many fiber subscribers do we need to add annually to fully offset the decline. So this is like what if modeling modeling the future type of a
[22:27] scenario the the question right you know within two seconds genie figured this out figure the math out the group need to add approximately 1.4 to 1.6 six million net new fiber subscriber annually and then um there are lots of
[22:45] insights and um some other strategy it gives you. Again this right at this moment you can feel the genie can serve as your trusted strate strategy as a strategy assistant to help you think and
[23:01] help you reason about the future. Okay. Um this is Chun. trend is just one side of the the um the factor. What about network network performance and network outage will cost you customers as well.
[23:17] So let's take a quick look besides the copex and besides the network uptime one most well most of the critical metrics is the MTR the meantime to resolution that means how fast we fix
[23:32] things because there is always um always a factors out of our control. you have whether you have earthquake or you have a fiber cut and what the leadership can control is how fast we react to those um
[23:48] to those events. So MTI is the matrix that we care about here. And let's play this with AI with AI agent. You can see it's not only able to flag every single
[24:03] event fire. It also it it gives you an alert for sure, but that's that's been around for decades. After alert, the AI is able to explain the triage process in plain English and translate every events
[24:19] into the customer impact and the dollar impact. Okay. And then further can that the what AI is really good at is the root cause analysis um and can provide autonomous solution autonomous fix if
[24:37] it's you know applicable or if needed they will just recommend hey we need to have a few technicians send it to the tower and fix the network. So every every events every network outage they can recover by itself autonomously is
[24:54] the cost you don't spend and the customer you don't lose. So okay so far this is the power of genie it brings the insight automatically to you without you even you know um doing that much of the that much of a thinking
[25:10] or drilling. So now with flip you ask um for CFO you can consolidate the frequent ask questions into a tab like this again we ask a very the exact the same
[25:26] question why business while line operating at a loss. Um what Genie is very good at is it's not only giving you the one analysis but also see it automatically write three more questions for you. The drill type of a question is
[25:42] slice and dice your data. Um I want to deep dive. I want to break down order by and so on. So that you can zoom into your data to do to do further reasoning. What if is modeling the future. You want your strategy team to say hey what if I
[25:57] increase the fiber investment and so on right um and why goes even one layer deeper than what if is the root cause analysis. So this type of a question after questions deep dive Genie is able to to provide to you and bring it to you
[26:15] just one click away instead of you need to submit a ticket you need to have a strategy meeting with your team and it it takes weeks. Um I think um pretty much with this it conclude the genie demo. I hope that
[26:32] inspire you as much as inspire myself. Um we have architecture here. Feel free to drop by our booth. I can explain how we build this type of genie demo. Um the data behind this is TM forum seat model align. So it's um you're probably
[26:49] familiar with the data model, but yeah, thank you. Let me hand over to my colleague Scott to um to go further.
[27:09] Let me get this set up here. Okay. All right. Let me go to Jamie here. I'm going to start a new chat. Okay. All right. Can we see this? Am I? That's good. Okay. Um, so what we're looking at here, and I might zoom in a little bit
[27:26] on the uh screen if I can. a little bit better. Okay. All right. So, what we're looking at here is a um a live simulation of a network that's
[27:42] built entirely in data bricks. Um this has towers, links, telemetry, all simulated and streaming within data bicks. Um, and when one of these sites fail on on screen here, it's thousands of customers losing service, SLA penalties. Um, oftentimes more than one
[27:59] truck roll required to to resolve the issue. Um, and a lot of times you're also having your engineers touch multiple different systems to obtain all the data points required um to kind of synthesize the full scope of of the error and the of the outage and of the
[28:16] impact. Um so in this demo what I'm what we're showing is data bricks bridging that gap and closing the loop serving as the data intelligence plane um for an autonomous network level four that in terms of uh TM forums definitions. Um on
[28:32] the left you can see we've got um an actual stream of data. You can see it's bringing about seven it should around 10 rows a second um of telemetry just to kind of illustrate the live the live data and it's healthy data right now. So the PM counters are showing positive
[28:47] values. You know what we would expect for a healthy network. Um what I'm going to do is inject degraded data into this stream. And so what's happening now actually in data bricks is that the the generator is actually going from healthy values healthy signal to noise ratios
[29:04] down to bad values essentially. Um, and what that's going to do is trigger a threshold that is going to have um that that is the agent will recognize that it needs to now solve something. So as we go here, the agent is detecting and
[29:20] classifying those signals as um as a particular error that then needs to be solved. So it classified it as a visual like RF path issue. Um maps the full operational picture in seconds, the site, the equipment, the customers. Um, and really importantly, it uses open-
[29:37] source hotel models to embed and retrieve both the equipment and the runbooks. So, manuals that we've given it to determine what um how how to take and triage this action are all done with AT&T open source models. Um, quantifies the business impact in dollars. This is
[29:53] all happening kind of behind the scenes here and it's going to give us a readout. Um, and it uses again those same hotel models to retrieve the right policy and the right procedure. And then the agent has a set of tools that it's been given um to remediate certain certain activities. So the LLM isn't
[30:09] actually touching the network ever. It's just selecting for instance an MCP server that will um that will actually suggest that the the change to the um service management orchestration layer. And as you can see here, the towers in blue are um essentially getting a
[30:27] coverage compensation order. And so while that tower is down, I know that went somewhat quick. Um, I'll bring this up over here. So, while that tower was down, um, essentially what it's doing is adding adding 3dBs to the surrounding sites to maintain the coverage on on
[30:42] that area. And I'll try to zoom in here a little bit more. You can stencil too much. Um, and so here you can see that it was able to create essentially um create an error based off of the off of the telemetry it
[30:58] was seeing. um it grounded all of its responses within actual runbooks that we've embedded um in data bricks gives a root cause analysis and confidence score for it again based in runbooks and vendor vendor manuals um it deres the
[31:15] users that have been impacted the MR that might be at risk due to that site going down um cross references it with several Service Now tickets and then before it does anything it it simulates its action against the digital twin um and make sure that the the neighbors have the capacity to absorb
[31:33] and ensure that they're not going to degrade and you know ruin the customer experience for that subset of customers. So it's essentially making sure that its action is safe. Um and then it'll it'll submit that policy guidance to the service management orchestration layer um and again base it in the runbooks and
[31:51] documentation. And in this case, it'll also dispatch a technician based off of where they are, geoloccation, the equipment on the truck, and the skill sets that that um technician has. So that we try to dispatch the right technician when there is a technician that needs to be dispatched. Um all of
[32:08] this kind of in in the model as a service light that we were talking about earlier. The left side here is done with um triaged essentially with claude haiku and claude sonnet depending on the speed that's required and the the need for intelligence within the model. And on
[32:23] the right we corroborate that response with AT&T's open-source hotel models that have been trained on telecom specific corpuses to validate that response. And so this could either be um what you're looking at here could be something of a report that gets sent to
[32:38] a network engineer. So it helps them essentially determine um issues without having to touch multiple systems. Or this could be written um as we'll see here as policy or intent to um to the to the SML layer. Um and all of this is all
[32:55] of it's governed um within Unity Catalog. we can interact with the data in plain in plain text using Genie and every single action that the agent takes is is traceable in ML flow. Um so first here I can I have a genie on the side
[33:11] and I can just say what is the latest event you have access to and what it should do is bring up the coverage compensation order that we just saw happen. Everything that you see on screen is written to Lakebase and then synced to the to Lakehouse so that we can use um use it for analytics. And
[33:29] I'll also show once this comes back in live demo. Um see so here that there's the event that we just saw. It's uh at this tower that we were just looking at. Used the it used the tool coverage compensation. Um it's
[33:45] now completed. Um and it had a fast response. And here we can see a little bit. I know it might be a bit of a challenge on screen but we can see that um activity here and if we go and open a new genie page so this is what's behind this is data bricks genie right this is
[34:01] behind the scenes this is the UI um and I could I can say show me all coverage comp compensation orders and so we get visibility into the action of the agent both in natural language and then as we'll see here in a second
[34:18] through every step of the MLFlow trace. Um, and it really just allows you to interact with that data and um, have the visibility, the traceability, and the auditability that's that's required to trust an agent within a network environment. Um, see if this comes back.
[34:40] Apologies on the slow internet a little bit. still thinking. All right. So then here we have all of the coverage compensation orders that you can see that have been
[34:56] submitted. Um this is demo so there's a few gaps in the days here if you will. Um but you can really see everything. You can see the neighbor sites that it's selected um for that DB increase. You can actually see the the the the intent that was passed. Um and uh we can look
[35:15] generally. Let me see here. We should be able to see usually the RPO and um the the subscribers impacted. Um so that Genie is is great. It it'll allow you to have that natural um interaction with
[35:30] your with your data. But then if we go back to our lake, go back to the lakehouse here and we can see in our traces, we can actually look at every single action that's been taken by that by that agent and how it's how they're
[35:48] evaluated in terms of did they site the right documents? Were they did they follow the right procedures? Was it a safe safe response? And did they pick the right tool? Um, so if we look at this latest trace, it hasn't been scored yet. Takes a few minutes to sync. Um so if we look at this prior trace and we
[36:04] look at the details you can really within within data bricks you can see the full chain of activity on the left all the MCP tools it's calling the runbooks it's retrieving um how it's using the hotel LLMs um and here you can really see just about every single thing
[36:21] that that agent did and and why it did it um so you get that full observability visibility into it actually retrieving the site info the device status, the the manuals, the the context it's assembling, um the runbooks it's retrieving, you can see the chunks that
[36:37] it that it chooses. Um you can see all of the MCPs that it's leveraging um and really get to very detailed information as to how um how the the the agent came to its conclusion and and what it actually did within your network. And so
[36:54] these would be the runbooks that are that it pulled the chunks from. Um and really get that full visibility into the triage. Um so this is really trying to show how again you can close the loop within data bricks. how we can go from, you know, a
[37:10] somewhat of a fragmented system to one that is is is essentially a closed loop with full visibility into things like your cost, um, your tokens, the latency, and and really all the steps and and the judges here. As we can see it, I've
[37:27] essentially created a couple judges that are somewhat arbitrary, but says that they have to be cited in a particular way. I want it to retrieve the documentation. And here we can see how that you know how this agent has performed against these judges. Um so it passed the tool tool choice correctness.
[37:44] It it selected the right uh tool to triage this type of alarm. Um and it cited all of its documentation the way I was expecting it to. Um so it passed both of those and it also passed the safety um the automated LM you know
[37:59] safety scoring. So within one platform, we're able to go from a network error to um directly into you know a triage scenario where we're trying to remove humans in the loop is as as responsibly as we can um and and try to achieve that
[38:17] autonomous network level four. So that is uh that is the uh that is the demo here. Thanks for coming. Um, if you want to see any more of this, uh, we will be
[38:34] down, all this stuff is running down at our, uh, industry solution booth. Feel free to stop by. We'll be happy to walk you through it in more detail. Thanks.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.