Ep 5: How AI is transforming revenue management - with Garth Lund and Charles Pierre

Advanced AI and machine-learning techniques are transforming how airlines optimize their revenue. Wiremind has been at the forefront of this shift since 2014. 

In this episode, Charles Pierre, CTO and Co-founder of Wiremind, joins Garth Lund to explore the future of revenue management. This deep dive covers AI’s evolution in revenue management, what airline RM can learn from other industries, and the potential of agentic AI to improve pricing decisions. 

(0:57) Intro to Wiremind

(03:31) What airline RM can learn from the rail and bus industries 

(07:17) What is best-in-class use of contextual data for pricing decisions

(10:10) Forecasting and optimization in Wiremind

(14:28) Introducing agentic RM 

(21:30) Combatting the black box and other challenges for AI in RM

(27:11) Bringing RM to ancillary pricing

Published by The Engine Cowl 

Transcript

SPEAKER_01 0:00

Hello, welcome to the Airline Tech Podcast. I'm Garth Lund. In this episode, I'll be speaking with Charles Pierre, CTO and co-founder of WireMind. WireMind's been developing revenue management systems across a number of different industries, including the airline business, for the past 10 years or so. And I think in recent years we've seen more and more disruption coming into the RM space with more airlines now starting to use techniques from AI or machine learning to optimize their pricing. And that's exactly what WireMind does. From my perspective, I think they're one of the companies which is closest to the technology frontier. And as you'll hear from Charles, they're super passionate about what they're doing. Hi Charles, how are you? Welcome. Hello, I'm I'm I'm fine. Thank you for having me. Great. So I think we'll get right into it. Um so first question: can you give us a brief introduction to WireMind, how you got started, what are the pain points you're looking to solve?

SPEAKER_00 0:57

Yeah, sure. Uh so my name is Charles, I'm one of the co-founder and the CTO of WireMind, uh company that I founded more than 10 years ago now. Uh and what we do is that we solve revenue management uh issue in rail, bus, and airline industries. Um, so we are around 200 people now. We have been profitable from start from the from day one and we are growing uh organically. And what we bring to the airline is uh a few things. First would be the advanced machine learning for forecasting and elasticity. Um and now we are adding on top of this uh uh algorithm algorithm art uh some agentic approach, as this is something that is going to transform a lot of industries. But not only about mathematics, we also solve a few operational challenges, um like a real time. So it's very important for us to have all our algorithms work in real time to be able to continuously re-optimize uh the network of our clients. Uh and we bring also an unified platform where not only you can optimize and uh pilot the seat inventory but also ancillary and overbooking, since they share the same core problem in terms of modelisation. And and the technical DNA of the company is to be uh a very engineering uh company. So we have more than 80% of our employees that are uh from uh in general background. Uh we are based in Paris and we really uh invest a lot into uh innovating and bringing uh the next uh the next approach to all of our customers. Sure.

SPEAKER_01 2:24

So as far as I know, a lot of the co-founders or early employees of the company came from an RM background as well, um, either in airline or in other industries. Was there something that um the team kind of was looking for on the market and they thought was missing, or what was the kind of drive to to set up WireMind?

SPEAKER_00 2:43

What we saw when we started the company with Skolin, which you might associate, is that you had either a very uh black box mathematical uh algorithms that would solve the problem, or you had a very uh let's say uh naive or not intelligent tool that were essentially just uh newicks, and we wanted to bring both of them into the same tools. And also what was uh missing in every tool that we uh are able to use was the the the fact that it was not developed by RM people. So this is something that is very important to us. Everyone that is working on the product side have a revenue management background, so they have used all of the RM tools uh available on the market, they know the weakness and they know what they wanted to have when they were analysts at the different companies that we are that they were working on.

SPEAKER_01 3:31

And so you mentioned there in the introduction that WireMind um is working across a number of different industries, uh rail, airline, bus, even uh ticketing as well, I believe. Which would you say is the most challenging uh industry for revenue management?

SPEAKER_00 3:48

I get this question a lot. I think actually it's not very about which one is the most challenging, it's more that they have very different problems. They also share some similarities, of course. Uh, what I can say is that uh on our three core vertical, which are bus, rail, and airline, for instance, bus, you have a very specific problem with high frequency, very small capacity, and many OND on a given bus. So this makes the modelization very hard because you don't have a lot of signal uh on which to learn. On the railway, it's more about uh doing revenue management that is not only optimizing revenue because they are under a lot of legal constraints. We are working with public clients, so they not only want to optimize the revenue but but they do it in a way that uh also favorize traffic, load factor, respect some constraint about the maximum price that they can set. Uh the fact that some ND has to be uh priced higher than short OND, everything uh that the legislator will impose to them has to be taken into account during the optimization phase. And of course the the airlines are very difficult in their own sense because the the demand is uh there is a high variance in the demand. The factors that change the demand are always fluctuating uh based on uh geopatical context. And the most different point with our other industry is the relentless competition that you see in airline that impacts very strongly the demand. So based on on your position compared to your competitors, the demand will not look alike at all. So this is something very specific to Airnet.

SPEAKER_01 5:15

Are there things that you think airlines can learn in terms of revenue management from the rail or bus industries?

SPEAKER_00 5:22

So first thing would be uh uh in in railway the demand materialize more in the last in the last few days of sales. So they are really good at uh analy uh at yielding the last day of sales, and they are very bullish on this, so they will uh make sure that there is enough capacity for the last three days in order to really get the most margin out of these days. When with our airline customer, we see more uh conservative approach where they try to fill the capacity before that that date or to leave a very few amount of sales in the last day. It makes sense on some OND, especially the long one, but on other OND that could behave like rail, it could be uh interesting to try a more risky approach on this part. Another point that is uh quite different is the battery steering. So when you work with railway, you know they have like uh all of the departures that are really close to each other, so they are very good at uh yielding multiple trains at the same time in order to ensure that there is a similarity in terms of price and nudging passengers to the shoulder trains in order to favorize the highest uh the highest price on the on the heart of the of the demand at 9pm and 9 a.m. for instance. And the last one is more technical and IT related is that rail operator uh they were quite uh far before, uh but now they are advanced, they are more advanced than some airline uh in the reservation system IT because they they did a switch a few years back to uh modern PSS or what they call inventory system in railway. And the PSS that we are working on with the railway, they offer much more capacity capabilities than the one that we are using on the airline side, especially on real time, push notification, uh continuous pricing. Those are some capabilities that we have already in the railway. And in the airline it's coming, but it's a bit slower. Okay, that's interesting.

SPEAKER_01 7:16

So in recent years, I think we've seen more and more contextual data used in airline RM compared to fair data, for example, is now uh very common as an input to pricing decisions. What do you think um the best airlines or best in class uh use of contextual data is right now, and where do you think that goes in three or five years' time?

SPEAKER_00 7:38

Uh good question. Uh for me, the biggest uh uh signal that you can use right now is uh the real-time booking data. So it's very important to have the optimization and the modeling works in real time to base its demand based on the last few sales that you made, because this is a kind of catch-all signals. So if you have a feature that you have not in your machine learning models, real time will give you the insight that something that is unexpected is happening and the model can react. So we have a this trend uh feature in the model that will almost immediately react to the fact that you have a highest or lowest or lower demand than than expected by the the model and it is able to auto-correct itself without learning all of the possible features. Then we have something that is uh that a lot of our customers are are are talking uh about is the look to book ratio. It's very interesting. Many people are actually looking at a flight on on a website on our set channels, but this is something that looks promising but is actually very hard to put in production because you need this uh real-time uh feed and API to get this data. And so far, to be honest, we don't have any customers that are able to give us that, but I see it a very promising uh uh feature to be used. And then uh I would say that in the future, what I what I see is that the large lang model like ChatGPT unlock a new kind of external data, uh mainly on the event and calendar data. So for now, what we are using is that it looks simple to have all of the holidays, school break, sport concerts inside some kind of calendar that the model can use, but in detail it's very hard to do because you have to do uh some cleaning by hand every time. It's very hard to get this data in a formatted or API way. And this is where ChatGPT can bring you something that is very valuable, that it can actually do this continuously. It will clean the data for you and create the structure output that you can use in the model. So this is something that we are investing quite a lot right now. We have uh one person working on this full time in order to bridge this gap and be able to have always up-to-date uh calendar data.

SPEAKER_01 9:45

You mentioned there the the use of the look-to-book ratio. Is that something you already see in other industries or it's new for rail, bus, or airline?

SPEAKER_00 9:53

Yeah, in railway it's easier because the the uh the way that you distribute uh your fares and and and booking is it's much more unified. So the railway they they master more their uh distribution channel, so they have this information for us to get.

SPEAKER_01 10:09

One of the more interesting elements I think of uh WireMind and the the Kaizen model is the the two-step optimization that you use for forecasting and optimization. Can you explain for us a bit more about how that works?

SPEAKER_00 10:23

Yes, uh so this is uh specifically the thing that I'm working uh almost all of my time at WireMind, so uh I'm kind of the expert of it. Uh so uh in our tool, the way that we do optimization is in a two-step phase. The first step would be demand modeling. So it's a very hard problem for two reasons. First one is the data scarcity. So you you only observe one price for every day before departure, basically, so you don't see all of the possible scenarios. And from this uh single price, you are supposed to learn what will happen at different prices. So it's it's basically a very hard problem because you don't have enough data compared to what you are trying to do. Um and so it's very important to use the maximum amount of data. And the second problem that arises quite quickly uh is the difference between correlation and causality. So let me explain a bit. Our uh customers they behave rationally. So in the past, when they have uh flights that have high demand, they were going to put high price because obviously they know that the demand is there. And on the opposite, when they have low demand, they will pull they would put low price to steer. Uh and that way, uh if you look at the data and you just give this data like this to a machine learning model, it will learn on correlation uh which is wrong. It will learn that if you increase the price, uh the demand will follow and increase. This is obviously wrong. It's something that is not common in machine learning problems where causality is not aligned with correlation. So to handle this, you have to use very specific uh machine learning framework called infer causal inference. This is something that is uh a branch of machine learning uh right now. It's a it's a research topic for a few teams around the world and it's it's quite difficult to fix. So the way that we modelize demand is actually very uh uh complex. So we use multiple models. The idea is to be able to really modelize the impact of price on the demand independent of all other co-founders. So we are going to basically learn different models. One where we learn the the demand without giving him the feature of the price, and then we are separately going to learn a demand that will only learn the impact on price on the demand. And then we combine all of these models to give you what we call the demand matrix. I cannot show you because uh it's a podcast, but in the application the the analyst can see what we expect the demand to look like for every price and for every day before departure for any given flight date. And then it becomes uh more of an algorithm problems where you are going to find the path inside the matrix that maximize the revenue under constraint. And the the constraints are both physical, so you cannot sell more synths than the the plane has, obviously, and also commercial. So for instance, uh some of our customers they don't want the the price to change every two hours. So there is a parameter that allows to respect some stabilities in terms of of time of booking. Uh, there is also uh an objective mix that allows the analyst to really set the goal between having a high load factor and high revenue because sometimes those goals they they don't align. Sometimes it's better to sell one seat at uh 1000 euro than ten seats at uh ninety euro. Uh all of this has to be taken into account in the algorithm to uh to really uh uh find the the path in the matrix that that uh represents the best uh solution uh for the commercial guidelines that the analyst has.

SPEAKER_01 13:46

Okay.

SPEAKER_00 13:47

And and on top of this, now we are working on the on the last layer, which is the agentic. So all of this is very mathematical. It's it's working quite well, but you see have a a few few cases where it doesn't behave like an analysis would and take decisions that are risky. This is where the agentic first it can catch all of these things. It's also easier for analysts to speak with an agent, to give him his commercial guideline, and then the agent will use uh the optimization and the algorithm like a tool to simulate different scenarios, all of possible paths, and find the path that is the best aligned with your interest. So it's really a blend of both techniques uh to have uh the full picture.

SPEAKER_01 14:28

So I guess the the the simplified output is the demand matrix um with the demand for each price point at each day to go, and you're looking to trace that optimal uh path. Once you layer in the uh capacity constraints on top of that, uh then it can give you kind of the optimized set of allocations at each at each day to go, right? Can you expand a bit on how you're developing um those agents to run on the estimated demand?

SPEAKER_00 14:58

We did an experiment internalys, so we took a lot of uh analyst decisions of the past uh and we tried to see if an agent could reproduce this decision. So we taught an uh an agent to think like an analyst. We give him uh you know like deal principle like okay, if you're above competition you should behave like this. If you are behind your reference curve, maybe you should like uh lower the price to get back to to the to the reference curve. All of this kind of uh you know like uh principles that an analyst have in mind when he's doing uh his work. And then uh we saw we we tried to see if the agent was able to do the same thing, the same decision as analyst did in the past, and and the results were quite good. So now uh we have an agent for which less than 7% of the time an analyst is not doesn't agree with the action of the of the agent. Um either the agent takes the right decision or it goes in the right direction, at least so it's very like the analyst cannot say no, it's completely wrong. Only 7% of the case you say okay, I will not do the same thing as the uh agent. And then when you have this, you have an in button independent system from the mathematical optimization, and this is where it becomes interesting is that you can use your independent system to check on the optimization. And when you do this, you actually uh see all of the cases where optimization and agent are are doesn't agree together. And on this case, 80% of the time the agent was right. So it's it was actually yielding a situation where we could improve the optimization or more likely the parameters of the optimization were not allowing him to take the right decision.

SPEAKER_01 16:34

So if you look at your customers currently, what is the the balance of the pricing decisions in terms of what's done by the model or by the agents and what's done by uh a human analyst typically?

SPEAKER_00 16:46

Yeah, for now uh the agent is not in production, so it's zero percent for the agent, and and we don't want to release it like an independent system. We really want uh to have something on top of the algorithm and the mathematical modeling because both has to work together. So the agent can steer the optimization. This is the way that we want to do it. The agent can take into account all of the possible paths that the optimization will put forward to take the right decision. So this is the missing piece that we are working on. And and to answer your question about optimization, it depends on customers. For some customers, 100% of their perimeter in other optimization, as long as we have historical data for them to modelize the demand. And the way that we work in the system is that optimization is not an adversary to other tools. So in in most flights, an analyst can take a decision and the optimization will start from this decision or correct itself based on this decision. And it's the same for a very important module that we have, which is the business rule engine. So all of these modules are designed to work together, and a good flight is going to be a multiple uh decision making from different sources, not the single optimization.

SPEAKER_01 17:54

So in the end, you'll have essentially, I guess, three elements. One is the optimization models, one is the agent, and one is the the human analyst for for steering.

SPEAKER_00 18:03

Most of the time they steer the optimization. So they will modify the parameters of the optimization to apply the commercial strategies that they want, for instance to improve generosity of the of the optimization, but they will not change the price themselves.

SPEAKER_01 18:16

So the analyst should be essentially managing the system, and then the system should manage the flight. Exactly. Okay. From my perspective, one of the I think most interesting selling points of WireMind is that you do enable airlines to um start with this business rule, you know, more traditional business rule logic, and you can layer on top of that um the AI or ML functionalities. How does that journey typically go for an airline or even a bus company or a rail company that's used to kind of more traditional techniques and they're moving towards more AI capabilities?

SPEAKER_00 18:52

Good question. Uh for most of our customers, there is an apprentice phase where they will indeed work with the system without optimization fully activated. So no airlines are going to activate the optimization from zero to 100% of their perimeter in one go. They all do it in a progressive rollout fashion. So they will start with uh some route and increase this perimeter along the way as they as they as they learn how to steer the optimization. But uh before that they they do it with business rules, so it's very easy to reproduce the business rule of a previous system in our system. The the UAX and UI is made it is made uh really simple to write business rule, and actually we also have an agent that write business rule, so the barrier to write the business rule is very low in the system. Uh and then the business rules they don't disappear. The business rule for us is a way to control the optimization too. So we have two kinds of business rules. Business rules that can set up alerts, a way that the analyst will be alerted about something uh happening that is unusual in a flight, some like a flight that is very well behind his reference or is to is uh is going too too fast. Or also business rules that would change the optimization parameter. So for instance, uh the stability that I was talking uh earlier is something that can change along the way. So for some uh many airlines they want a high stability at the beginning of the of the booking curve where you don't want to change price every two days, but on the last day of sales, on the last 30 days, maybe you are you want to allow the system to change the price every six hours. So this is something that is very uh easy to achieve using a business rule. Um and um and uh one last thing is that uh in terms of of journey, so we have done some proof of concept where we were against uh another revenue management system, and for all the route we were the new system, the other route we are the old system, and even without doing any full implementation or true uh on the analyst, just by setting the system with full optimization on, we already saw uh ROI. So for us, of course, it's uh the very uh powerful case on the fact that the system without even uh supervision is already good enough, but the the true uh the the true capability only appears when you really learn how to use the system and to steer it. And and the model will also improve itself as it goes on, uh, because most of the time when we implement for a first time a customer, we don't we get historical data, but it is not as clean as we can have once the system is in production where we are going to gather all of the data, all of the signals that we process.

SPEAKER_01 21:30

A common critique of um AI-based revenue management systems is that the optimization can become a bit of a black box for the analysts. How would you respond to that potential concern that some revenue management teams might have?

SPEAKER_00 21:45

Actually, uh we work with uh quite uh passionate people, so they always want to see uh how the model behaves and to be able to run their own uh you know, like dashboard and forecast accuracy measurement. So we have a very uh glass box. approach on this. All of the accuracy measurement you can get it from the application either at a micro level flight date. We are giving you the error on the past 10% of the of the capacity cells. The extraction we have API for the customer to extract all of the data, including forecast and observed booking so they can also run their own validation. And we are on the progress of rolling out something that is I think will be very useful is a dashboard to see the error at a more macro level to really understand what is the price elasticity of the model, how it behaves on each on each route, is it good at what when at one day before departure is it good or not? And this is actually the dashboard that we use internally or our data science team use internally to validate new models. So it's a way for us to align our validation with the customer because now we are going to share the same dashboard and the same same matrix to be sure that during the training phase of the machine learning we are actually using the KPI that matter the most for the customer. And this is very important because if you a model can be trained with a lot of different targets and and loss and and the and the idea is to use the loss that matter the most for the customer.

SPEAKER_01 23:14

We touched earlier on um a few of the challenges for training AI models within revenue management. In particular you mentioned the data scarcity um and also the you know this problem of um concluding the wrong correlation when you're looking at the data between demand and and price level. I think another of the challenges is that the there's no you there's no correct solution um to measure against when you're training the model. How do you deal with that um with that challenge?

SPEAKER_00 23:46

For the data scarcity the idea is to use the maximum amount of data that you have and the most valuable data is always the latest one. So what is happening currently on on the flight that just departed or even on the flight that are not yet departed is the the sign the signal that you want to use the most in the training. So this is not uh technically it's not that hard but in terms of modeling it you have to be careful because you don't want to overlearn on some data and then having a model that will not generalize well. So this is something that we are doing but you have to be very careful about the mathematical truth behind it. And you can uh for us the way that we do it is that we have automated retraining pipeline the in a way that machine learning can retrain a few times in a year to make sure that we always use the latest data that we have in in the database and to make sure that we cross validate in order to make sure that the model does generalize and he has not learned something very specific on the coming months and which make it drift a lot for the the month that he has not seen. And then on the ground truth about on the revenue management, yes it's very hard to so validating the forecast is simple. So you have like mathematical metrics like the amount of errors you made the amount of booking that a flight did and the amount that you predicted this is a very easy way to to see if your prediction was right. On the optimization part it's much harder because you cannot say okay the price asset was indeed the optimal one and on this the best way to measure is to have an independent system like a the one I described before an agent that could reproduce what an analyst uh takes as as action and to use this as a judge of the optimization. So this is the way forward for us is to use this LLM agent to challenge the optimization and and spot the the situation where you can take a decision that are a bit uh different. Another way is to make sure that the optimization doesn't recommend prices that are very different from the past. You don't want to recommend always the the the price that you saw in the historical data because you don't do it optimization by doing this, you just reproduce the past. But you don't want because it's very risky to to recommend something that is twice higher in terms of price. So you have to to assess the elasticity of the model by making sure that when you give him no constraint at all the price he will select is still in the range of probable price compared to last year.

SPEAKER_01 26:23

And are those constraints something that you would typically expect the analyst to input or that's something that's already built into the model?

SPEAKER_00 26:30

No this is something that we uh we fully built on on our own. But then there is like this endover where after every training we have a product owner team on the data science part that will present the result to the to the analyst or the the airlines and we'll be very uh forward about it. So we'll tell you on which route we think that the model is good enough be below 20% of error and on which route we think it's not good enough because we don't have any any historical data because the network has changed a lot there is multiple reasons but the important part is that the analyst knows where it can behave good or bad and to make their own decision based on the result of the model.

SPEAKER_01 27:10

Okay. So I believe you've recently been developing a model um for optimizing ancillary pricing. Can you explain a bit how that interacts with the ticket pricing and also how that then helps to optimize total revenue across the business?

SPEAKER_00 27:26

So ancillary pricing for us it's very uh interesting and uh it brings quite a lot also on on seat optimization for for multiple reasons. So first the problem is is kind of the same it's just that you don't have this capacity constraints. So the optimization part is easier but the modalization part is as much as difficult as the seat part. What you want to learn is the attachment rate which is the probability that someone is going to buy an ancillary at a certain price. So first you have the same inference causal framework because there is the price notion and you have to be able to also learn a price sensitivity. But then something that is you don't have in seats and I would like to be able to bring it to seat afterward is the the dynamic pricing part because attachment rates they really behave differently if your booking has three passengers, if it has only single passenger, if the origin destination is uh something like six hour flight, is it during an holiday in which case most likely you want uh a luggage so all of these uh parameters of of every booking is very important to give you the attachment rate and so we were able to learn uh also the segment uh that makes it that set this attachment rate. So the way the ancillary uh modelisation works is that first you are going to segment the demand based on on those kind of features and once you have all of these segments you are going to price each segment independently. So in a way it's it's the same as if you had a seat for which you could sell at multiple price depending on the people buying it. And I think it's the future for revenue management and and this should be bring uh to to the to the seat inventory at some point. And lastly also because the PSS for ancillary is much more modern you don't have this bucket system. So you can work on a continuous pricing way and and this is for revenue management and algorithm it's uh it's it's very powerful because now you can put your price just below competition you can use every price exactly uh to get the right uh the right revenue um so all of this makes ancillary uh modelization a bit harder but also more advanced than seat and it's very uh knowledgeable and and uh valuable to learn all of this on ancillary to to be able then to bring to bring it back to revenue of the seat. So for now we have done this uh with one of our customers and we are live with them. They use exactly the same interface as the seat to steer the ancillary so they have access to uh the optimization to the business rule engine to be able to steer the price of ancillary from the business rule is very powerful especially in the beginning because uh if you go from uh you know like fixed price ancillary to dynamic pricing you don't have the historical data to learn price sensitivity. So you have a phase where the business rule is going to be the only way for you to try out different prices and to acquire this data that then we are going to be able to feed to the model. And the step further that we are developing but we are not quite there yet is to be uh is doing optimization to optimize both at the same time because sometimes you you want the seat optimization to understand that it's better to sell at a lower price because the people that are it's going to buy will have a higher chance of buying an ancillary. But for now we we we do separate optimization. We are learning uh how auxiliary optimization works, how it behaves and and in the coming months we are going to try to to merge both but uh for now it's separate.

SPEAKER_01 30:55

Yeah I think that'll be super interesting if you can merge the the decision making between the two. Okay one final question if there was one thing that you wish Airlines knew about WireMind or the Kaisen model um what would that be?

SPEAKER_00 31:07

I think it would be that uh Airlines has to understand that when you work with the RMS system you not only choose a software but you also choose partnerships you choose a company that is able to to to grow with you and to accompany your your your growth and and your new development in terms of commercial product. So this is what Wire was built on. We have this very strong culture of innovation of engineering culture we don't want to to just develop a product and then don't touch it for the next 10 years. We are really in the forefront of of uh not only trying things like uh continuous pricing ancillary but also uh agentic something that is going to change uh the way that we work uh all the way that we interact with our application uh is going to change in the next year and and you have to choose a partner that understands that and have the technical capabilities to transform the application to go uh this way to not be stuck in the past when all of our competition or the rest of the industry will have moved forward. And lastly especially in RMS the scientific depth uh not death but depth of the company is very important. So the core of the science you have to have a a scientific course someone that like a research team but you don't want to do research in like a laboratory with no implementation whatsoever in the production. So bringing around the table the expert, the RM analyst, the customer, the product guy that is able to like design a system that you know in terms of uh of design is an Apple level product because now we all expect to have systems that looks like a very well refined the science and of course the technical and software engineering to make it in real time to make it scalable to have an architecture that doesn't break every two hours is the strength of WireMind to have all of these people around the same table and and to build a product together for the airline and and with them as partner. Thanks.

SPEAKER_01 33:06

So I think this was a super interesting discussion especially for any uh RM geeks or AI geeks out there. If anyone's feeling inspired about wiremind what's the best way for them to get in touch with you or with the the team?

SPEAKER_00 33:16

Yeah you can connect to the website Wireman.io and there is a form for you to to get touch with us uh myself available you can find me on LinkedIn my name is Charles Pierce uh and and if you are more into like a science and mathematical stuff uh know that we are uh publishing some papers we have published one uh about press celestially last year and we are expecting to to work uh on a few coming months on next paper on on diffusion model something that is coming from the large language model that we are trying to see if it it can fit actually the the RM problem.

SPEAKER_01 33:50

Okay very interesting well thanks a lot for your time Charles