Databricks Powers Real-Time Energy Analytics at Octopus

0:00so I guess our first question actually is can you explain to everyone what is octopus energy yeah good question so octopus energy now is quite a big group of companies so I often get that question what is octopus energy and what what do we do we are eight years old as a group of companies and we started really as a as an energy retail business so we the aim when we started was to

0:22disrupt the energy retail sector we saw a sector that was really not serving customers in the UK so very bad customer service outcomes some very some some slightly lasty pricing behaviors and just generally not a great sort of customer experience but we also saw a sector that was basically the delivery route for Net Zero so where we see it is that to get to Net Zero you have to get customers on board so you can't do this

0:45behind the scenes we tried it we and we've done an amazing job particularly in the UK things like wind turbines and solar panels and customers kind of didn't really know that was happening you know they could see them in the in like off the coast but they didn't know necessarily other than a little bit of extra cost on their bills but now we're at the phase of Net Zero transition where we have to engage customers we

1:07have to change the cars they drive we have to change how we heat our homes from sort of burning gas in our homes to using things like heat pumps and that needs a delivery arm and so we saw energy retail as the sort of way to deliver that so having a relationship with the customer and then being able to take them on that Journey from maybe buying some renewable energy to having a electric vehicle and a heat p pump and

1:30then letting us control that electric vehicle and heat pump so that it uses the most green energy at the most green times and so effectively all all we do as a group of company is aimed at delivering that Net Zero transition from a customer perspective and so it kind of breaks down into energy retail so we sell energy we buy energy and we sell it to customers in I think eight countries now around the world we have a

1:52technology arm because called Kraken and the aim of Kraken is basically to be that platform the operating system for a a mod energy utility and we license that to I think 50% of the UK market now and uh large chunk of the Australian Market making big inroads in Japan so and that the idea there is that you you can't change an industry without changing like the sort of the software that it runs on

2:15and the software was quite old and not really fit for purpose so we built our M platform to do that we have a generation arm where we build renewable generation so we go out and invest in win turbance you know the hard stuff and then we have a Services business where we go out and actually like literally go to people's homes in our octopus branded Vans and go and change meters and St heat pumps and electric vehicles and we're trying to do

2:37that globally and you tell us a bit more like what is the role of data in the transition to n zero to run an energy system you need to have loads of data you need to understand how much energy is being used at any time you have to forecast that you need to make sure you're dispatching the amount of energy you need at that point in time and to do that requires a lot of data and even in the old world where we had you know a

2:59few 100 power stations that National Grid could dispatch and then demand was pretty predictable because everybody behaved the same that was still like quite data driven in the new world where you've got like all these electric vehicles which are doing different things and maybe being dispatched against price signals and we've got and we have to balance the grid like much lower in the grid because there's so much more demand at the end nodes there's so much more data required so we

3:22ingest vast amounts of data both in our retail businesses Our Generation businesses flexibility business and obviously Within in our software business and it's crucial to making great decisions what would you say are like the main sources of data what are the largest and most complex ones you have to deal with there's like width and depth the widest data sources the sort of like the data sources with the most

3:44tables and columns and like the most sort of higher complexity come from our our platform Kraken because it's keeping it's like a full stack energy platform it does everything from speaking to the meter all the way through to transcribing calls and doing clever AI stuff on it it's it literally every part of a customer journey is captured by by Kraken and so there's hundreds and hundreds of tables to ingest and just

4:08getting all those tables into a data platform and making use of them is a challenge for our data engineering teams and then the Deep tables are things like smart meter data so a smart meter spits out in the UK 48 readings so one every half an hour in Australia one every five minutes and so if you're running millions of customers on platform it's you know it's not huge data compared to some of the stuff that some of the big Tech Giants do but it is quite a lot of

4:31data and getting that process aside and then soon we're starting to see that being overtaken by Telemetry data so data that's coming off us are are like wind turbines or data that's coming off the electric vehicles and telling us like you know I've plugged in I plugged out I want to charge I want I need to charge now so yeah we have we've had to you know we rep platforms on a while ago because we saw that change coming like

4:54just tsunami of data that we had to handle you mentioned earlier you wanted people using the most green energy the most green times is that a fixed time the most green times or is that when your wind turbines are saying like oh it's really windy it used they do say that it used to be a fixed time in the sense that you could predict that during 4 or 7 p.m. that's

5:17when the like most people come home and like like statistics Works in large numbers at at a grid level like you you kind of knew that like every day the big peak was going to be there and the and that generally meant you're going to have to drive all the way up through your generation stack so from nuclear and then you're going to have to like bring on all of your thermal so like your coal and then your all of your whatever Renewables on at that point and

5:39then maybe your gas diesel it's all the way up so you get to your dirtiest stack by the in the peak hours now it's actually much harder to predict when is going to be the greenest half an hour of the day so you kind of know what the dirtiest time is it's always going to be that Peak period but the cleanest time is harder to predict because it might be in the middle of the night when there's loads of wind on in the summer it might be in the middle of the day

6:01when there's loads of solar on the system a lot of installed solar in the UK and I'm talking a bit of UK but like it's very different in say Japan or in Australia where they've got this enormous solar load um so it's it's hard to predict being able to predict that and being able to move demand into that is extremely valuable for the customer but it's also valuable for the for the utility how responsive is that let's say

6:24there's like an emergency government broadcast everyone turns on their television at once no one coming how quickly do can you go you know if it was just the wind turbines going how does that work to go let's say you need to go right up to the dirty sources how quickly can that happen and how does that occur so and just to be clear like octopus like in each country there's a

6:46system operator and the Energy System operator in the UK that's National Grid and it's their job to dispatch all this it's our job to provide them with the tools to do that and what I mean by that is like tradition their tools have been pull up the biggest power station and and turn it up uh and you and they'll kind of plan for that in advance but they'll also have a bunch of spinning Reserve which they can just

7:09turn on and turn up they've also got frequency response which is like stuff which they don't even have to dispatch it just react and in what we what we realized is that like customers have a huge potential to flex and they have even more potential to flex when they've got an electric vehicle which is a giant battery sitting on their driveway and they have even more potential to flex when they have an electric heat pump and the home which is effectively a

7:31ginormous thermal battery so what we're trying to do is find C great customer propositions that can that can incentivize customers to flex and then ways to be able to package that flex and sell it back to the system operator what do you mean by Flex Flex is a basically the ability to move load or sometimes destroy load and generally it's moving load so it's basically

7:53saying there's a bunch of load coming a peak time we want to move that to the middle of the how would you do that and traditionally the system operator wouldn't move load so they would say demand is inflexible was the Paradigm the thinking Paradigm a few years ago and generation is flexible so you reshape your generation to match demand in the world we see coming and and what

8:16the world we're designing our platforms and our businesses for generation is harder to predict and more intermittent because it's coming from Renewables and demand is much more flexible and and you can dispatch it and so we can effectively we can take like the idea is we can take like a bunch of EVs and move that those that EV load to the best greenest time when there's the most power in the system and it's best for

8:39the system operator whilst also fulfilling the needs of the customer and the needs of the customer might be I just need my car charge by 6 a.m. so I don't care when I charge it is it midnight is it 900 p.m. 10 p.m. whatever that definitely makes sense there must be a massive Network effect it and how hard do you find it to change consumer habits to try and get everyone doing that flexible model we were originally told that customers

9:01won't Flex you know you can't persuade a customer to change when they do their washing or their dishwashing you can't pers the customer to charge their car we looked at like businesses which were ahead or or sectors which were ahead of the like digitalization curve and realize that customers really do Flex so like it's a bit of a cliche but Uber is an amazing analogy because they realize that if you give customers a very clear

9:25price signal with a nice UI and a very clear reward signal for for like effectively a surge The Surge charge for not taking a taxi in a in a specific period or or having to pay a premium for for a busy period they will firstly like the taxes will come to that area and that time and also the customers will find other alternative routes and so we

9:47I I guess like like we saw that and realized that like like a really simple incentive structure and a really nice UI with the two things that made made this work and so we've come up with a few options we have a we smart tariffs where the price of energy you pay depends on literally depends on like would not change their demand patterns and do you think that started

10:09to change the perception in the industry I think the industry needed showing so we took a lot of risk we basically launched smart tariffs which where the pricing formula like for our agile one we were sort of artificially at some point sending it below zero to do plunge pricing because we wanted to see what customers would do when prices were negative and now we have a system where prices do go negative quite a lot but we

10:32were you know the formula we put in meant that that happened a bit more and so yeah we I think we felt like we needed to show the system there were lots of people by the way who are already you know in the industry saying you know this could happen so we kind of helped I think Galvanize that and the nice thing is that the system operators are really on board now the network operators are really on board now I think not just in the UK but in all in

10:54like many countries that people they see the value of this and for some like saving sessions which is a national incentive it's called the DFS demand flexibility service I think we now have loads and loads of suppliers taking part and not just loads of suppliers they're also now companies that aren't even energy suppliers that that have like jumped on the bandwagon because they see the value in this business model and

11:16they doing the same thing just like I think D like can you tell us a bit more about your team like you have a great Mission and your lead of data and how many people do you have in your team how you keep them productive focused and what are they doing very good question and defining team is sometimes hard I I think of it more like a community or a function now and the way we organize is uh what I

11:39would call like platform pods and community so platform is a central data platform team who provide all of the infrastructure that our data stuff runs on and that is tens of Engineers and a mix of platform engineers and and data Engineers so people building like actual Cloud infrastructure versus people working in Databricks and building pipelines and deploying them and then we

12:02have PODS of people for each of our business unit and in the bigger business units pods per like department so in the retail business we might have a finance and Commercial pod and that's a group of data scientists and data analysts mixed discipline who are embedded within the business unit and their job is to go after there like specific business problems in that area and we've and then the community is like the thing I try

12:25and knit together which is like how do you make everyone feel part of like a dat a strategy and a and a group of data professionals who are really like awesome and ahead of and doing great stuff and actually it's a fun place to work so we do that through like meetings and channels and all that stuff it's really interesting particularly because o has been so fast growing and how do you manage that as the team grows because it's great when you're all kind

12:47of like a small close group of people but as a company scales very rapidly do you encounter any challenges and trying to keep that culture yeah so so it's worth saying I was the data team for a year or so so I was the only data person in a startup of like 20 people and now I sort of dodged your question about how many people are we but I think when we get together like globally across all of our business units are like data

13:09function is like 50 to right now and it just started with you and it started with me yeah and so the way I see it is like think is in phases so like phase one is like startup phase and all you care about is solving like immediate burning business problems you don't care about what tools you're using you don't care about like

13:32whether you have development Frameworks and team structures when when it was just me there was no point in having development Frameworks and team structures all you care about is like the CEO has a growth problem how do you solve that immediately in the quickest way possible then you go into like scale up phase where maybe we had a team of like six or seven people at that point we started to focus a bit more tooling

13:54because you like you can't scale analytics Beyond a few people unless you're using like you know shared repositories Version Control and also the data you know you can't like if everybody's running different SQL queries on their machine like what's the source of Truth so we started to think started to think about like how do we maintain like consistent data pipelines

14:16and then I'd say we're now at Enterprise scale and and we're worried about like how do you organize the team to deliver across like a vast array of different business units how do we make sure we have the right Tools in place that scale with the data volumes so that's when we started things at Databricks how do you maintain a project with hundreds of data developers so our DBT projects have like

14:39over 100 contributors our Databricks SQL warehouses have over 600 queriers so like and it's like it's a layering problem so like your inner layer is your data team your data platform team your next layer is your data pods and then your layer out is like all these people who are doing amazing things with data like hundreds of them but I can't see everything they're doing and nor should

15:01I and so how do you give them the tools the training to get what to like do the right thing and Empower them without constraining them but also without them doing creating a whole bunch of chaos those are the that's what I think is the Enterprise challenge now and I guess data governance came into question a lot our approach has been or DBT like tooling has helped us a

15:25lot so DBT helped us with massive data governance and then you catalog is helping us a lot now little shout out catalog well the thing the big thing for me about data governance was like when we were small everybody like I used to get like endless emails from companies trying to sell me data governance and I was like you know this was like sort of startup stage I was like what I don't really know exactly what data governance is and but in my head it was like

15:48something Draconian which meant like not doing things and then I sort of started when we started to hit like data governance challenges you know which is the right table to use for this specific problem or then like actually data governance is really an enabler it's like a positive thing not a negative thing it's it's being able to see all of your data clearly having one view of it having owners so people know who to go

16:10to when there's a problem having like access like really strong access controls so that you know that it's secure by Design rather than secure by like somebody scrambling around to keep it secure and once we hit that we're like okay what's the tooling that we need to do that and and I think yeah as I said DBT is great because it maintains your all of your transformation in one place and unity catalog is also really awesome because it means it's like it's

16:32bringing together the sort of like user layer and the sort of asset layer so the users who's using it and the assets what are the tables in one platform which really helps I find the example of data governance really interesting in terms of like how quickly you scaled and lesson you learn and my follow-up question would be if you were to look back now and you had like one grain of wisdom for someone in a similar like

16:55high growth first person the data team in a startup what would you say that you would have done differently or you would have considered differently I think there's some stuff which we probably should have got to earlier so things like I wish I'd known about DBT earlier because it would have just helped us and I wish we'd we went down a strategy of using like sort of presto Athena type tooling initially

17:18which lacked which was great as like very high performance very low cost scalable compute but it lacked that like data governance element or like just the organizational like organizing your data sources they've added a lot of stuff to it recently which bridges that Gap but I definitely think if we'd introduced those a bit earlier we might have like

17:41staved off some of the issues I also think it's like it's almost easier and harder now as a startup because there's really well established tooling now you know people call the modern data stack it's quite easy to like I know I'm safe buying Databricks or or big query or or snowflake or or even like sort of duck DB if you're small plus a DBT or a SQL Mash plus a sort of like visualization

18:05layer I think when we were starting some of that stuff was around but it was like less Pro it was less clear like this is exactly what you should choose and use although there's so many more tools now that it's probably like slightly Choice Paradox people coming into the space yeah and that's really interesting actually like these modern data platforms are as you say like they they're much more mature and but I think

18:27one thing that your team really put them on actually is like open source Technologies I was wondering how do you kind of consider both of those and like when you look to solve a problem obviously you can get a data platform but at the same time you generally also want to innovate in your own ways and how do you when you're considering a tool for a solution like how do you take your account open source versus propriety data platform so my thesis

18:49having been burnt a few times on choosing tools is the engineers and the analysts and scientists are they're going to be the ones working on it it's not going to be me because I'm answering emails and doing meetings and and trying to organize and so generally the best tools we've had have been the ones chosen by the team and so and often promoted by the team they like hey I think we should use this DBT was a good example of that streamlet is another

19:11good example when the team they looked at a few things I asked them to look at a few things but they said this is the one we really like stre let's use that so yeah I think like picking the tools that the teams like and often Engineers if you Engineers quite like open source because they can see under Ro so yeah we're like often looking to balance what the engineers want to use and then obviously you have you know is it

19:34Enterprise grade can you know does it have all the Octor Integrations and security stuff that you need but generally you can't go too far wrong if you like let the people using the tools pick the tools because ultimately they're working with them day to day very Democratic I think it's actually quite unusual I don't of my team will agree with whether I've lived up to that but we definitely AFF this

19:56part and can I tell us a bit more about data quality because I'm assuming if you're working on use cases such as like Supply demand signaling having data quality issues can lead to very wrong results so what is your approach how you do this the problems with data quality I actually think tend to rise typically in like what I'd call like the primary

20:18layer so the source system that's often where data quality is the biggest issue and then that will roll all the way through to your like your data analytics there so all the way from ingestion through if it's captured poorly or or like is just wrong in that L you've got youve you've already got an uph challenge we're quite lucky in that a lot of particularly in the retail

20:40business a lot of the data we ingest comes from a system that that Kraken that we as a as a group of businesses have built and it's really well architected and so generally doesn't have as like loads and loads of messy data in it it's actually very clean and well modeled so that makes life a bit easier I think then sort of take again taking that approach of like load it all into the lake and then use DBT to

21:03incrementally clean it up I think the danger is where you have like hidden clean up layers so if you're like you've got a clean up layer sort of somewhere in the ingestion level that the analyst can't see they're going to be like well why is like in the source system it looks like this but in just it looks like this why is that I can't understand that so being able to see like where you make your changes um and and often what

21:25I say to the analyst is like understand the business business process that's creating your data so like if you're looking at say sign up data literally go yourself and sign up that product go through every page yourself so you and then look at where the data arrives in the database so you fully understand how that data is captured and then if there are data quality issues sometimes data quality issues are perceived data quality issues because they haven't

21:48understood the sign up process that's generating the data where it's not a perceived problem it's a real problem then address that either at either with the source system itself or put in a fix somewhere up the stream that's that's well documented I think what's really interesting to me is particularly when I read the news and when I consider my own like energy suppliers is actually Smart Meters because there's a lot around them actually being a bit unreliable and you

22:11touched on how you basically actually have Kraken under the hood to kind of help you ensure that data is actually high quality when it arrives in the lake and I was wondering like how do you deal with challenges such as smart meters and maybe you can also elaborate a bit more on what does Kraken exactly do yeah so I I'll talk through what Kraken does and then I'll address this because I should have answered data quality with Smart

22:34Meters because that is a very good example of where we don't control the data capture and it can be really challenging so Kraken Kraken is three parts it's Kraken customer Kraken field and Kraken Flex so craen customer is the retail platform it basically allows an energy supplier or utility in in a regulated market to run their whole systems so literally everything from capturing meter readings all the way

22:56through to building those meter readings customers get in touch all like all of the customer contact like any sort of apis that are exposed to sort of front ends all handled by and then fle I won't touch on flex and field right now I think for the smart meter question they're less relevant with smart meters in most countries the supplier doesn't look after the smart meter so the smart

23:18meter is typically put in place either by the network owner so like in London it's ukpn in Paris it's enodis like the company that looks after the networks sometimes look after the smart meter that's what happens in um countries like France I think in Japan and they collect the data from the smart meter and they might collect that centrally so they might put it all in one big database and

23:41then the supplier can ask for it and if they if they own the meter point they look if they look after that me they'll get the data in the UK it's different in that the supplier technically owns the meter but the infrastructure to get the data from the meter to a sort of central Hub is run by a company called the DCC so it's centrally run not by the supplier not by

24:04the network operator but by this company and the construct in UK is like every meter look like the data lives on the meter and you just request it nobody centrally stores that data whereas in other countries you pull the data in you clean it up and you make a nice Central database that you can then Source out so we've got this like weird system where in the UK we're relying on these like not very reliable databases at the end

24:27of a not very reliable work to keep track of all our smart metata we then request it from the DCC we put it in our in our data store and then we process it and we use a lot of Databricks to process it because it's a lot of data and that's where we then will fill in data quality so we'll do anomaly detection we'll do like infill reads where we haven't got the data we'll go and request reads where we have where we've got a gap where we think the data is on the meter but we hasn't come to us

24:50and that is a really hard problem and it's a classic case of like if you don't own the data capture you're kind of beholden to the data capture system and you have to do data quality solving down Downstream rather than out the source that's a really hard problem actually I was wondering do you use any kind of AI and machine learning to handle that or I wondering like what exactly you do

25:13in that case it's it's actually nothing hugely fancy so there is there is plenty of like AI built into into cracken in other places but I think uh for the most part it's like simple an anomaly detection like you know is it is the read higher than much higher than the typical read flat meter you can call that machine learning if you want it's a sort of comparison of a reading to A

25:36mean and then infill is like you're looking at like some sort of like baselining of previous periods to fill infill holes so it doesn't it's not like rocket science it's actually the data engineering which is much harder in the smart met processing space yeah I think that's really interesting because I guess I also come from like a data science background and I think that as a data scientist with a good team of data Engineers you could be quite best

25:58because you don't understand just how much effort goes into the ingestion process and the cleaning process I guess like when you're talking about you have these business units they have these pods and that's is that why you have generally like like different teams or having like a diverse range different skill sets to kind of help if you have a data scientist they obviously need a data engineer there to help clean up in the first place so we tend to want we

26:20tend to have most of the engineers in the platform team because we want to do most like the ingestion clean up in that team and not have like because that process is kind of the same for particular the ingestion bit and the sort of staging is the same regardless of which department you're in you could be in electric vehicle department or you could be in the building department of the retail business it's always the same like stuff so we try and centralize that

26:42the pods should then be what I'd call like full stack data professionals I think is a better term uh data scientists and data analyst is quite a fuzzy Spectrum so I quite like sort of full stack data professionals and what I mean by that is like they should be able to listen to the business owner find a problem figure out that problem and they should be able to build the data assets that they need

27:04to answer that problem and that data asset might be a whole new DBT table it might be a new dashboard it might be a machine learning model that you deploy it might just be a graph that you do and stick in a Google doc because you need to do it quickly like whatever it is that solves the problem and so the idea is the pods should be like pretty have like high agency and be very close to

27:26the business problems and and the platform team should be building tooling that enables that so you don't have when I look at some data organizations they have the problem of like it versus business so the data analysts and data Engineers live in the tools atbt in an Ideal World we should have like senior business leader speaking to data analysts and they solve the whole problem and only when there's something

27:50like new ingestion or the platforms like you like something's broken should we have to then go to the platform team I had a question about like you know interaction of business users with a data team and what are some of the good practices of communicating results from the analysis to the business users and do you believe in data democratization so the business users can interact with

28:14the data directly Maybe by leveraging generative Ai and all this new techniques yeah so I'm a very firm believer in data democratization Al I think it's something that gets banded around and does and it's hard to achieve and doesn't get like I don't think we fully nailed it yet but lots of people talk about it and then it actually doesn't happen I told you earlier that we had over 600 Databricks platform

28:37users and a and a data team of you know 50 to 100 people depending on how you count it that means that we're like six to one leveraged in terms of the number of people interacting with data at like a Databricks platform level so that might be dashboarding or it might be using the SQL UI versus like the people like the data professionals so yeah loads of

29:00people are basically using that data and so I think we've sort of democratized that and I think gen will hopefully start to bridge that gap between I have a question and I need to write a query the key enabler there is you have to make sure your data is in good shape and well documented because if it's not thei is going to write like a great query

29:23against the wrong table Yeah and then the communication thing I'd love to talk about that one CU I think it's really interesting and something maybe that like falls down particular amongst Junior data people I see that come in they typically will focused on tooling so they're like I'm great I can write the best DBT I'm really good at SQL like nailed git I'm really good at python but

29:45they haven't focused on the other two ends of the of the data Journey which is like can you synthesize the problem can you figure out like what actually the business is asking for and can you communicate your results really effectively so communication for me is like the mag trick that opens up is to like imagine you're receiving the email rather than writing the email so if I'm receiving that email like what

30:09does that like just look at it from that eye like that access doesn't make any sense or like uh that's an ambiguous statement or I haven't put enough information in here to actually describe what I did or where I got the data from and as soon as you start flipping it and saying what would my what would the person I'm sending this to ask as soon as you get that empathy with the person you're doing the data for the quicker

30:31you get really good at communicating again like that's the fast growing company side of things I just EP problem will get increasingly hard as you continue to grow and like how do you handle this when you have for example you have like you mentioned you have Kraken field Kraken customer all these different units do you do like knowledge sharing between business units to find out how you're solving similar problems and you or do you been find this problem

30:53at the moment actually we definitely do I think we have like things like sort of Central data science standups and data like data help channels which kind of helps so people can see what's coming in something we did recently partic we find this problem particularly where we've got retail businesses in lots of countries and they're all like the UK has six million

31:15customers and it's very mature eny retail business our us business is like thousands of customers and is like not as mature yet they're going for a different journey and I don't want them to solve problems like we solve problems because they because we have to solve problems a bit slower because there's a lot more data and a lot more like customers in play and like you you can't

31:37take as many risks I don't think as when you're like super startup but one one one challenge that I was put to me was like I don't know what the other regions are doing and I don't know like they surely someone's like solved this problem somewhere else and we tried a meeting but the meeting was everyone said their thing and nobody really like everybody was like this is what I'm working on and kind of no you know like you don't really listen when everyone's

31:59just like listing out what they're working on I think we've pushed it Forward recently what I've done is asked for a shared road map so in every like domain area so write down what your country is doing and then people can go and cherry pick from that we're going to publish that to the whole the whole retail data organization so they can go and be like hey I'm looking at I know Germany are looking at that thing I might going to speak to them so I think of it's it's like it's about bilateral

32:21collaboration encouraging bilateral collaboration rather than encouraging centralized collaboration the more you can get people to speak each other and know what they're doing better rather than trying to sensually push out like hey you work with you or try and do a one-sized fital solution across every region yeah is I find it really interesting because I guess like how do you also encourage that also from even

32:44like an incentive structure because I guess coming frombe some larger Enterprises it can be maybe almost competitive in large companies actually when you have different teams different units and how do you ensure that collaboration is got is always positive and you're always encouraging like these positive behaviors that's a that's a very good question I think generally if you

33:06hire we have less of an issue where we're like where we've got lots of competitive tension between business particularly like you know the German retail business the UK really wants to sck grow do well so we're like we want to help them so I think we like don't really have as much a problem if you hire like people with high eqs they'll tend to resolve conflict rather than

33:29like create conflict and I the I think the challenge comes when you're competing for resource so say like three countries want something from the different things from the data platform team and then you need engineering managers again with like high eqs and the ability to like juggle but also find the like quickest solution or like the common solution like hey what could we do here that like could solve

33:51everybody's problem yeah that totally makes sense you you clearly put like a huge emphasis actually on not just technical skills but all Al the soft skills which is really interesting because I feel like in the past Tech was very much focused around the half skills and it's much more of like a a Modern Trend to towards this one thing you did touch on the earlier was a little bit on generative AI I just love to dig into that because it's it's so topical right now we not

34:13ask I guess like even just starting off like how do you see it on the hype curve do you think it's overhype do you think there's real business value immediately you can unlock what's your kind of opinions on it so there's I mean it's definitely I think it's like simultaneously sort of Shing as cat Vibes of like it's sort of simultaneously overhyped and extremely valuable from a business perspective

34:35like it is definitely adding lots of I can see adding lots of value across our organization but I also see a lot of hype out there and being able to cut through the hype and figure out like what are the tools that really add value and what is somebody who's created a business put a wrapper around open an open AI API and is charging for that like that business I think then the way to solve that is to is to try and work

34:58like as as low down the stack as possible so build stuff on top of like foundational model apis as opposed to buying lots and lots of wrapped Services because the models are moving so fast are you able to share like what kind of business challenges even at a high level do you think that it's most appropriate to solve right now because there's a lot of discussion around you know issues of hallucinations and the other quirky

35:20parts of these models and how do you kind of balance the benefits and the risks so I see it in two I sort of spit it out when I talk about J there's the enabling piece which is like almost like an it or a tech Ops type challenge like gen is going to be in every application on everyone's machines and it's just it's literally like Excel when Excel

35:42came out it's like it's a new tool that anyone can use for any it's like a new SP army knife and then there's like gen in product where you embed your gen into something you're building and serving to customers it's all on the sort of like enabling generic side it's obviously super powerful for as a as like co-pilot like we lot of GitHub co-pilot a lot of people use it just for day-to-day

36:05helping them answer questions and doing their job particularly if they're writing a lot of code because it's it's awesome at coding I use it a lot when I'm hacking stuff together and then on the product side we speak to a lot of customers like hundreds of thousands of customers a week and so as a platform you know Kraken is enabling U to speak to customers and to having great

36:27Tooling in that makes a lot of sense and so we've there's a few products that the Kraken team built into it that help our energy Specialists our customer service teams speak to customers and solve their problems quicker so one is like summarization which is awesome so we can summarize a call like a seven minute call with a customer if if if an energy specialist was coming back to that call to say what happened with that customer

36:49they'd have to listen to seven minutes of audio we just transcribe it and summarize it and so they can read to like 30 seconds of reading gets in that they get six and a half minutes back and then we can do the same with long threads so if they've like long long messaging threads we can do the same like what's going on we do sort of message generation assistance so where the energy specialist is like can you

37:11just write a quick response confirming this thing it will take all of the context all of our knowledge bases and help the en and basically generate a message in the right tone of voice for the specialist always check before it goes out um and now we're moving into what I'd call like autom at response and action and so that is like taking a custom message saying actually I know

37:33what this is I know how to solve it generating a response again not automated we will check that it's okay and generating an action plan so it might be processes Meer reading and issue a bill and and here are the two buttons to do that and so the Ken team are building as well so I think it's got like huge potential and I think voice I've seen a few demos recently where vo is like the latency on the conversation

37:55is coming down a lot and the ability to handle like interruptions and just like the general complexity of a human conversation is coming down so like voice is coming fast and God only knows what that means for customer service companies fastic and um the about checkin response that was generated by the model what is your approach then do you have like a human in a loop are you

38:18looking into llm as a judge like because I think human in the loop is great to have but it doesn't scale too much so what so we're using human in the loop the idea is the Gen is if I take octopus is like is like found ourselves on like amazing customer service and we in the UK have been like I think seven years in

38:41a row or six years in a row the witch the only wit recommended energy provider we really like love and are passionate about doing like going above and beyond for customers and so we'll always want humans to help customers with like their difficult problems because often like an energy problem energy is like one of the biggest bills you pay and and often an energy problem

39:03is can be extremely distressing or or hard like if you go into into debt you can't pay your bills you know if your prepay meters cut you off and you no longer have heat like that's a really difficult personal problem that you don't want to speak to a robot for you want a human to solve it and you want to do it like amazing empathy and then and maybe go above and beyond like you a

39:25little a little gesture to like make them feel loved and like and appreciated and robots are not you know AIS are not going to be able to do that maybe they'll be able to sort of emulate that but they won't be able to emulate it in a way that's sort of like real and like generous so the idea is we're building AI geni tools that make our energy Specialists more superum and when we

39:47started Kracken that was the whole point of Kraken before gen the Gen Trend the idea was like they have everything at their fingertips to be able to answer a customer question so in a tradition when you called up a traditional energy supplier you would get like I've got a I've got a payment issue and they go okay let me transfer you to the payments team and it actually turned out that payments issue was a billing issue so you get kicked onto the to the billing

40:09team and like you've spoken to three people maybe taking like an hour and maybe no one solved your problem because they've kicked it over to the back office team who have to action the thing later on and so the whole idea with Kraken was like it's a One-Stop shop the energy specialist is empowered to solve your whole problem because the platform does everything and that so that was a case of like using technology to make an

40:31energy specialist super human and we think gen is just the next layer on top of that there will be some stuff we can ultimate but we still want that like human human interaction that makes sense I've never thought about it and it's actually like you know we have some other service providers and being of kicked from as you said billing team to back office can be so frustrating so and yeah great to

40:56hear that your challenges approach very interesting because it's like it's the Paradigm of like designing your customer Journey around your systems which makes like if you start a startup now you never design your customer Journey around your your systems you design your customer journey in a way that is like the best for the customer and then you build your systems to deliver that but Legacy businesses for whatever reason design their customer Journey around

41:19their Source systems which is just like the put in the car for the horse but actually one thing I did want to ask a little bit was like your journey to the field like to data in particular because I think you actually have an energy background but I was wondering if you kind of share like how you ended up at octopus and just a little bit around that yeah so I'm like uh maybe I call myself like a sort of data mongle I I

41:43haven't come from a sort of like algorithmic PhD machine learning background um nor have I come from a software engineering background i' I've sort of taught myself those domains um I worked in energy trading for uh BP for for a few years and there I kind of fell in love with like the Energy System I think it's like unbelievably fascinating this like complex web of infrastructure

42:07markets people that delivers what we take for granted but like keeping a grid on and keeping everyone's cars filled up all the time at quite a low cost is an amazing feat of like both physical engineering economic engineering and like human engineering and I really like love being part of that so I loved energy I want I wanted to be sort of closer to the action within DP so closer

42:31to sort of like the hardware so I then moved into reservoir modeling so I was like doing computational modeling to figure out how many how you get oil out the ground and there people were doing like you have to run like lots of simulations so it's like lots of montea simulations to figure out like because there's so much uncertainty because you're trying to work out what's going on like like thousands of meters under

42:55the ground no one really knows all you've done is like poke a few holes in it and then like that's your information you poked a few holes you've like like sent some seismic sound down and looked at the reflections and that is your understanding and you then trying to plan like exactly how to deploy hundreds of millions of pounds worth of capital to go extract that oil so you end up running loads and loads of simulations and in that role people were like doing

43:19them one by one you have to like a simulation is effectively a computer program that you run in in a Linux machine and they were doing them one by one and I was like why why don't we just like batch these up I saw I started hacking together Python scripts and then I was like hacking together sort of like databas type like results formats and then sticking the tools on top I was like this is fun I like I'm quite

43:41enjoying like building these data but I had no idea like what else was out there I also had a sneaking suspicion that like the oil and gas sector probably wasn't like morally the place I wanted to spend the rest of my career and so I started looking around for jobs in Tech and clean tech and I found octopus and I meet the founders and they told me this compelling story about kind of a lot of stuff I described which is like using technology to both drive a customer Le

44:04transition and also like do it in a way that's like great for the customer and you and so s took a pun left an enormous company to join a 20 person company built all of our like pricing and commercial models so we like did a bunch of work figuring out where should be priced that was the first question answer like where should we pric our energy to be competitive but also

44:27cost effective and then from there built out a lot of analytic stack and then hired a load of people who are like way more Technical and better than me and have built a team from there so that's my journey iing is like so I would say I come from more of a business and Energy background I spend my a lot of my evenings trying to backfill knowledge My My My Philosophy is like I should be

44:49able to roughly do what like most people in the team or at least understand it so like certainly there bits of like the kuber stack that we use I have like very different clue but at least like I'll have a crack at like running a little mini Cube machine on my machine once just to understand the concept because I think if you get to a place in an organization where your leaders don't understand the day-to-day jobs of what's

45:12going on you end up with like really disjointed decision making so I'll like annoy my team by like taking on little bits of work or like asking them stupid technical questions but that allows me to M like remain techn technologically sharp I think when you're talked about energy trading I just remembered the movie andron the smartest guys in the room I don't know have you seen it I

45:34read the book yeah well I guess it's good to see like companies like octopus energy and data leaders like yourself who have like the opposite principles from that crowd the energy sector is often painted it's like e the evil versus the and and obviously N1 there was like massive corporate failings and some like really nefarious Behavior d by some very bad incentive

45:58structures that led to like some really bad outcomes but in the energy sector generally it is full regardless if you're working on like the oil and gas side a lot of the people I work with BP are like amazing super smart want to do the right thing or like the clean energy side like the energy sector is full of like really incredible passionate people and I think the key is

46:21for us we see the goal is like to lead try and Lead that sector to do the right thing and sometimes that means you know there's a lot of received wisdom in the energy sector as well because it's been around for so long and and it works like does work like stuff you get like we have petrol cars and lights in this room but being able to show that there is a different way to do it like you can you know customers can be paid to turn down

46:43we don't have to spend millions of pounds like turning down a turning up a nuclear power station that's the key and technolog is like obv Central I also have a final question I remember from earlier it's a hard one though harder than the mron one it's actually easier um it's it's back to the topic we can't avoid which is Gen um because I find it really interesting when you're describing about having a human in the

47:05loop system because it's perhaps particularly topical given the recent case with Canada air and the refund policy and I guess one one thing I'd like to understand is although you have a human in the loop they can sometimes perhaps get fatigued or overly confident if they're just for example checking if the response is appropriate or not and how are you going to handle that in the context of perhaps if custom about a

47:27refund policy how do you handle that auto message generation yeah like it's a problem there's some amazing Stories coming I like the cisler one about the chatbot which was just a sort of thin wrap around open Ai and you could just use it as a cheap open AI um yeah there's stuff you have to guard against and and definitely like the like fatigue or the lack of checking is a is a

47:49really important one and then you know that can Cascade into gen making up your company policies for you one of the things one of the engineers came up with was really smart was this idea of like edit distance checking so we check the edit distance between the message we this the energy specialist ends up sending and the message that gets generated and so if you systematically

48:11look at that you'll find the energy Specialists who are not making any meaningful changes to the messages but also like this stuff you catch this stuff fast like like Greg is amazing in that he reads a lot of customer emails and gets a lot of escalations and answers them all on on social media Greg sorry the CEO of octopus and he and and that

48:33culture of like right at the very top level still interacting with customers means you catch this stuff so like the management team Optimus have an amazing nose for when something's going wrong because they're speaking to customers on a daily basis and so if you have systemic issues which we defend against as long as you catch them fast and then Rectify them that's that's really cool

48:56I'll say Thank you very much then for listening to the first episode of the digital na podcast and for our guest David for being our first guest yeah yes DAV thanks so much thank you for having

49:25me e