Instant Predictive Analytics on Your Own Data Using Terno AI

Key Takeaways

Hi everyone. Can you hear me? Yeah. Feel free to use the chat to respond. I'm waiting for the panelist to join. This is Glen join me. Hi everyone. Um, welcome to this session. Let me start with sharing my screen. Okay. All right. All right. So wonderful to have uh such a great audience today. Uh we have people from all walks of life and many of yeah from from various different countries. The the the topic of the webinar is that how to get the instant predictive analytics on your own data. So essentially the core idea is that how can you you know automatically generate the predictive analytics on your data. So Turno does exactly that. So we'll be walking through Turno and we'll walk you through like how it works and how can you build your own reports on your own data. Right? So and we'll also talk about uh right various uh use cases. We'll go we dig deeper into the use cases. I am Sandep Giri. I am founder of turno AI and uh with me are Rohan Glenn Tir and Kajinder. Rohan will be de demonstrating how to do the forecasting of the sales and then uh Glenn will be demonstrating the product recommendation system that he built using Turnoi and Tir will be talking about customer retention analytics how he was able to achieve the the analytics on customer retention using kerno and Gazinder will be talking about the customer segmentation based on the sales data. All right. So, so what exactly is this the brief this is briefly the agenda. We'll talk briefly about how it works, how what turno is and how it works and then we'll also talk about the various case studies. Feel free to ask questions at any point of time. Let me check if uh right okay if uh the there are any questions so far there are no questions and there's one question uh related to is the recording of the session made available later um this uh recording yes will be put on our YouTube channel okay and we will also send you an email but yeah don't don't drop from the call just because of that utilize this call to get answers to your questions. Okay. Also, yes. All right. Next. Uh okay. All right. All right. Just give me a moment. I'll make my teammates as okay so that they can answer questions in the chat. All right. A question from D is I thought there would be a local model which could help here without using AAI. Is this possible? We'll talk about that. We'll talk about that. Okay. Very soon we'll talk about it. And so the the reason why people use the local model is because they want to do the analytics on their own local data. So we'll talk about how you can achieve that. How you can achieve that using turno AI. Okay. All right. Any questions? Any other questions before we begin? So, what exactly is Turno AI? I hope you guys can hear me. >> All right. >> Yes, you're able. >> Okay. Okay. Just give me a second. Okay. So, what exactly is Turno AI? Turn is your AI data scientist. It essentially behaves like a data scientist and uh it I mean you can ask it questions, you can give it a task to do. It'll do all of that. It automates the data extraction, processing, analysis, visualization and model development. It is based on the task. The agent delivers results. Right? So whatever is your task, whatever is your ask, whatever question that you're asking based on that the agent delivers result that could be in the form of tables, charts, graphics or it could be text, PDF, images or even creating custom files such as you can ask it to write a code and build an app. So you can easily share the results with your team or schedule them to be delivered via email. So uh basically whatever results you get whether it's a PDF whether it's a whole conversation the graphs you can ask it to share the results with your team right and you can basically uh whatever the intermediate resources are there whether there are data sets models that can be used by your entire team once that has been built also. It it is kind of really intelligent in the sense that when you interact with it, when you tell it once about your data, it learns from that. So it learns with you and all of that knowledge stays inside your organization. So the way it works is that the user user whether uploads a file or gives a along with the question you can you can ask it a question that here is my file in the form of CSV and it can ask that it can answer that. So you can attach attach the the qu you can attach the file to the question and then that on top of it you on the other hand you have your data sources such as the data sources could be data links CRM ERPs MCPS like there are ERPs enterprise resource planning software such as ODO SAP CRM software such as Salesforce and so on and then there then there are these MCPs the MCPS are almost every organization has now every every software whether it's softify or anything they all have provided their MCP so you can connect the MCP uh of Shopify to turno and then when you ask it a question that how many orders have been shipped uh which were uh how many on an average how many delay has been in shipping of my Um you basically can ask it to generate that kind of report and so on or or you can you can essentially uh you know build a model whereby you can detect how many of the orders are fake or they they're going to be refunded versus uh uh uh not to be refunded. So those kind of uh you know um analytics those kind of models you can build using turno. So you can so user ask a question goes to turno. Turno writes code and and writes SQL and then talks to the data. So turno launches the code near to the data instead of you uploading the data to turno. the turno writes the code and sends that to be executed near the data because uh like the way it happens in case of chat GPD where you attach the data files to uh send to send to uh chat GPT and you get the results that's not the way um that's possible if your database is big enough you cannot afford the organizations cannot afford to send their data to chd or any other LLM. Therefore, turno writes code, writes the logic and executes it near the data and then once turno gets the results, it presents that to end user. Okay. So essentially um essentially turno does everything for you whether it's a Unix command it needs to do it needs to write SQLs it needs to write um it needs to uh write um Python code it'll it does everything starting from Python code SQL Linux command everything it does for you to get the results what you're looking for this is briefly how turn works so either like Usually what people organizations do is before starting with turno they connect all their data sources and then user asks a question it basically starts doing that and it keeps learning from your data from its own failures. It keeps learning on on the data. It learns from its failures. It learns from its successes. It learns from users questions. It learns from the data that you have in your organization. And it keeps on becoming smarter and smarter. And you can also explicitly teach the tribal knowledge to turno saying that that this was my old uh report. Can you learn from it? So it'll learn and figure out the the the local or expert knowledge of your organization from your past history. So this is briefly how it works. Now many of you might be asking how is it different from shared GPT? Because the moment you see Okay. So, so if you if you take a look at Turno, if you go to turno.ai, everyone, you can go to turno.ai, click on try now, and basically it'll ask you to fill a small form. Fill the form genuinely so that so that we can fine-tune your experience and then then wait for the approval. So my team will approve and once you get that by default we create one organization for you. You can create as many organizations as you want. For every organization you have a separate URL. Right? This the design is from the perspective of like everybody might have different organizations. They might be working for multiple organization because people are consultants. they have different um installations and so on and uh they want to also separate the data from one from another like for example we have created a demo data set the demo or so demo.app.turno Turno. So this is the hosted version. If you are in an organization, you can install Turno locally. We are also releasing a desktop version whereby you don't need to install you don't need to upload or connect your data to Turno. Instead, it will be local to you. Okay. So right now here we have connected these data sources. You can connect all your data sources whether data bras, snowflake, oracle, myql, bigquery, posgress and so on and you can just connect all of them like we have already connected odo erp and global data and demo sql and now you can start doing the the analytics. So the now here these are the these are the predefined queries we have uh done. Okay, these are the predefined queries where you can uh you can see it when you sign up these databases we by default we attach and then we can start asking questions. You can start asking questions such as find all customers who made more than one purchase in last three months for for each list total revenue, number of orders, salesperson and last order date presented as customer loyalty report and so on. So you can see that it is basically uh basically this will basically give you uh details about everything. Okay. So and here like here basically we have right so it has generated the the detailed report. If you see this is the customer uh loyalty report it has generated for me. Okay. And so on. All right. So you can essentially uh get it rolling. All right. So this is the one. Now moving on. Moving on. Um let's go back here. Okay. So essentially uh also on top of it basically the way it works is you can also do things like this. Right. I wanted to extract images from my PDF file. So I gave it my PDF. It separate it basically extracted the image. Okay. That's my passport. Okay. I'm just skipping over it. But yes, it does all of that. Okay, so I use it for all sorts of tasks, right? Like for example, here here I asked it analyze leads and won deals by campaign and and so on. It has done the decent job of preparing that analysis report. Okay, so it is doing all of that. You can see here. So revenue per lead is this much, right? This is the source like from newsletter newsletter we got this much revenue without the source we got this much revenue search engine we got this much revenue and like that so it does detailed analytics for you can connect all sources of your data whether it's Google analytics whether it's CRM whether it's um you know um ads account so from everywhere you can connect and then ask it to give you the detailed analytics. Okay, not only the normal analytics, you can also do the predictive analytics. Now with this, let's move on to the next question that how is it different from chat GPT. The difference lies in this diagram. Okay, difference lies in this diagram here. Basically if you go to chat GPT you will have to upload all of your data to chat GPT or any other LLM for that matter and then you'll get result in case of turno turno will connect will write code and run it on your data okay so that you don't have to upload data to turno okay now that also means that turno makes sure that it doesn't send your data uh to any any LLM instead it only needs to look at your metadata the information on the data further what turno can do is let's say you have really confidential data and you don't want to share it with any LLM what you can do is turn can build an app for the analytics on your dummy data and then you can deploy it deploy it in production on your real data So the app does not access any LLM instead app is just like an app. All right. So it follows that model. Now moving on. All right. Yeah. Turno AI is an agent. It does everything. It's it's the upper end of the agent whereby it uses something called a chain of thought processing. chain of thought processing. Chain of thought means it takes your question, breaks it down into intermediate tasks. Then for each task, it write code, it writes SQL, it uh uses search engine, it uses all sorts of techniques. If it needs to use third party tool such as let's say it needs to do OpenCV or any machine learning model or AI, it'll download those, install those, use that library to get that work done. It could be maybe you need to process video so it'll install the video processing library. Maybe you need to process the images so it'll install the image processing library. Maybe you need to use the scheduleuler and it schedule algorithms are like the public algorithms it'll download and start using. So it'll basically download and install all the libraries that are needed. Let's say audio processing, image synthesis, voice synthesis. It'll basically install those libraries and get the work done. Similarly, if you need to let's say prepare the presentation in PDF, it will install the the PDF library and get that done. Okay. All right. So, okay. So, basically it is exactly like a data scientist. Okay. Now GBT or plain LLMs none of them have built-in ability to connect your data understanding your database uh the way uh turno does it also turn provides something called a security layer which can be used to protect your database and to provide the au authorization on top of the existing rules. So all right. Okay. All right. So that's pretty much let me take some questions. Is there any question? Question from sur is what is the underlying LLM? That's a good question. Right now we are using 03 uh I think 03 mini and uh 03 mini from open AAI but we are compatible with almost all LLMs but as per our minimum requirement we would want the uh the models to be more uh what do you say uh more uh reasoning models. Okay. Any other questions? Okay. So, you can you can sign up and experience for yourself and try all kinds of questions for yourself. A question from Arouch is that can turn Aai be used for educational purposes specifically for students who want to be data scientists in future of course. So this is something that we are also doing very very very frequently. So the way people do is that right there are new ways of teaching uh teaching things right like um for example uh for example new tools is there so you need to utilize the tool to teach uh anybody right like for example in uh most of the exams calculator is allowed now right so similarly similarly we should teach using uh data science using turno we have already started that as as part of cloudex lab. So the way we do is that you take a problem statement ask to solve it and then observe what exactly the turno is doing and then and then you basically start nudging it the way you want it and that way you get to learn a lot about the data science in general. So yes, it can be used for learning learning data sciences and uh and it does a pretty good job there. A question is but when we use turno AI isn't it would be feedback to your model even on local data no we don't we don't your local knowledge whatever gathers that local knowledge is separated to your instance it doesn't come to us okay we will not be using it uh as part of our agreement with every organization Okay. How do you maintain the consistency of the output when you are using LLM? Usually LLM responses are not identical when same query is executed twice. That's right. That's right. That's why what we do is that's that's why what we do is that the the there is a there's a basically a lot of work goes in fine-tuning it fine-tuning an LLM for your organization. So for your organization we take all the queries questions that we have been done in past and let LLM learn from it. So after a while it comes up with very good knowledge and when you are and also at the in the beginning itself we when we work with large organization we have the the the design time whereby we build let's say 20 20 uh test cases and we keep on fine-tuning the prompt fine-tuning the organizational prompt until we get perfect for those 20 test cases. Okay, so we have we that's how we do that. That's how it basically works. Okay, and that's that's the important part and we have been working on making that consistent. So it's a difficult question but we probably our results have been extremely consistent and we are working on multiple levels on that. The first one that I just now described was that we maintain test cases for every organization or during implementation time we build these test cases and then these test cases run against your database uh very quickly and then check whether those test cases are all passing. Okay, between the two between the two uh calls to LLM the answer might differ but there can be multiple answers to the same question. Okay. So that's a important part that we have done. So LLMs are pro probabilistic in nature and that's what we are doing. That's what we doing. We're utilizing that knowledge that knowledge that that knowledge of probabilities that LLM has built in order to get the work done by the way of building a verification step by the way of building a building um you know um hard test cases. We have a hard test cases in the sense that when LLM does something it executes a code if the code is wrong it'll correct that code and and so on. So it basically we make it selfcorrect. We make it uh give it a verification step so that we can give it a feedback in real time and that's how we tame the animal called LLM in order to get the work done. Okay, very good questions. So do you have do you use rag embedding to store data from data sources? Not really. Not really. We do not. We have a very different strategy. Okay. So we do rag embeddings only for very few things. Okay. Because you cannot you cannot give answers on relational data just based on rack. For example, if you have a database table and you want to plot a nice chart about uh you know the what is the average a like let's say what is the average revenue average revenue that you got per customer over the years. This kind of question cannot be answered by a rag based system. Ragbased systems the way they work is that you have a question you match with the subsection of the data attach the two generate the answer right so basically basically that's that's the answer that's the idea so um rag doesn't work on data rag is good for text when you have PDFs and so on and you want to just generate an answer it's But when you are generating numerical values, it won't work on numeric data on data such as databases you it wouldn't make any sense. A question from Arun is is it uses rag or cg concept? It is basically a um we call it chain of thought. Chain of thought combined with a sandbox environment whereby it runs the python code. That's how it works. Okay. A question from uh right but isn't the desktop version coming would be fully local? No. The desktop version when we say desktop version uh means the execution environment where it'll run the code, it'll code and it'll save your file all of that will become local to your computer. Okay. It would still use the LLM our LLM. We have a we will have our own LLM like fine-tuned for such use cases. It will only interact with our LLM whereby it will only send the metadata not your data. So the desktop version will be the case where it'll run the code everything inside on at your desktop rather than on on turn AI. Okay. A question from Suresh is apart from ad hoc queries on my database is is turno bringing capabilities like PowerBI? Yes, that's a very interesting question sur. So what we are doing is we are building something called app services. If you uh look at my screen, I'll just show you something. Okay. So once you are done, let's say show me purchases made by customers in Canada. Oh, it's not yet up here. It's on it's on uh Turno app. We have a test environment. Okay. So this is not something for you. This is only um basically for internal testing. Okay. So here right uh you don't have to go here this is just for me this you can see that this is something called um basically the analytics okay so so here when we do something when we do something okay um right so when we do something or finish something we can ask it to create an app service. Okay, we can ask it to create an app service and I just now did something right. Yeah. So as soon as it finishes the work, it will a ask you to create an app service and then you can basically just launch that app service without without having uh to do any like having to do uh run the the LLM again. Okay. So it becomes just like it becomes just like your uh it becomes just like your um your PowerBI or anything. So you you spend time in building the app right you let let it let know build entire thing once you get good results you ask it you basically create an app ask it to create an app it will package everything into an app and then host it locally and then you can start using that so you don't have to essentially build the entire thing again and again I hope that makes sense uh Sur So yeah, go ahead. Uh, okay. Let me read your question. Yes, it will be published like a dashboard. Yeah, it'll publish like a dashboard and basically it'll be updated automatically. So that's the whole idea. Okay, a question from ANA is can it be used to pull up all sorts of data or are there any exceptions? It can be used for anything. It can be used for your custom services. It can be used. Okay. Uh yeah, a question from Anom is will there be a fees to use it? Yes, of course. We have to pay for OBDI. So, we have the fees. Fees is listed on Tanoi's page. So, you can take a look. All right. Now is the time for the demo. Okay. So first is uh by um Glenn. Glenn you want to share? I'll stop sharing my screen. Please go ahead. >> Oh thank you Sep for the wonderful uh explanation about turno. So we'll u just look at what uh I I'll present a practical case study that demonstrates the power and usability of turnoi. U to illustrate its capabilities. I'll walk through a real project or a real chat that we conducted using the big basket product data set. Right. Uh uh I hope my screen is uh shared. >> Hello Glenn. It's not shared yet. >> Is it now? >> Yeah. Yeah. >> Yes. So u okay uh I wanted to create a product recommendation system right so uh that was that was uh from the big basket product data set so presented entirely right so uh I gave a prompt just like a normal uh without a technical knowledge I wanted to create a product recommendation system right so I just I didn't even provide the data set I just provided the uh address of the data set uh as you can see uh my highlight so I also attached the JSON right the list of features merge CSV and I sort of give four to five task for it to do it so as sep explained the chain of thought right it's it explained what why what it is doing first it installed the Kaggle CLI and uh download and unzipped the uh data set from and uh uploaded it into the sandbox read all the JSON JSON files flattened and merged them so it gave the example uh explanation on what it is doing right so it when when it isn't successful in doing it reiterates and tries to do it again until it's successful So u right and it was successful in downloading the data set and uh sort of extract the uh data from the given kagasle URL. So it explained me what the URL contained. It located the file read the sheets u sort of uh selected renamed and uh understood the schema of the data set and saved the merge CSV and display displayed it first u rows. So if you can see the screen uh it gives it gave me a basic overview of what the data set contained right categories subcategories sub subcategories the link of the product E code image link brand SQ name SQ size right so it uh sort of created me a basic EDA process uh as in in technical terms it gave me the overview of the data set that I wanted to do it right It previewed my data set. It loaded the data set and it made me to easily verify the column in inspect rows. It also explained me the sum it also summarized the data set. Right? It uh explained me the abbreviations of the PS in the uh further responses and gave me some statistical uh useful statistics like uh price distribution, minimum, maximum, median, MRP. Right? So it gave me all the basic idea that I need to know about the data. Next is I asked it to handle the missing and duplicate values which is one of the uh pre-processing uh steps that we follow. So again it followed the chain of thought it explained what it is going to do and it did it. So it it it was pretty uh fast with the responses and it gave you the it gave us the overview of what it did in the previous uh from the previous uh prompt that was given. Next uh I wanted to build an interactive dashboard. I build I wanted to build an interactive dashboard on the explor given the data set right for business users we want what we see is what we believe right so uh visualizations come to come into play so I asked turno to build an interactive dashboard which it built using plot link right focus on four uh four questions major questions that it uh it perceive by itself which categories have the most products which subcategories dominate overall what are the top 10 brands by product count and distribution of MR price. So you need not even give what you want. It uh sort of basically understands through a business perspective on what you want or what you would like to see and it visualize it very beautifully uh for us to uh sort of understand u the visuals um the data in the visual format. So I was curious so I went to do further analysis. I wanted more detailed analysis. It computed the mean, median, mode, min, max and subcategory. Highlighted the most expensive ones as well. And the actionable insights help us to category managers to spot high value product segments quickly. Uh yeah, it sort of gave me highest. >> So there's a question that how is it interactive? You can you can actually just open that in a full screen then right full screen not that inside that uh yeah in yeah open that >> so I can just >> yeah you can see when you take the cursor it shows you the things okay you can interact with it there are the top there are ways to display things okay so it's actually uh unlike the static image you can see that when you zoom it it is responsive when you take the cursor on Right. You can see that you can interact with the data. Right. I hope that answers question. All right. Yeah. Go ahead. >> Right. Uh I hope that answered the question. Right. So u again I I wanted to deep dive it uh gave me the recommendations. The core feature of this was the recommendations. Turnup built a sophisticated end using TF recommendation and cosign similarity for any product. I can get five most similar items based on the name, brand and description, right? Instantly, right? Uh you see the example where mosquito repellent were highly relevant in the similar suggestions. So if you uh if I scroll down um so good night uh gold flash liquid vaporizer mosquito repellent were was the top five the top five recommendations based on the uh given data set right so further I wanted to analyze more I create I gave a simple command uh simple prompt like create a cluster plot of products based on the similarity and give me Clean and clear plot, right? Clean and clear plot. So, it was u it projected the products into two dimensions, clustered them using K means and produced fully interactive cluster plot as as Sundep explained a while ago. So, making it easier to understand the catalog dasa, right? It's uh and spot natural groupings or gaps in the data, right? So in conclusion, no. >> Can you take Can you take the full screen for this one? >> Yeah. Uh >> yeah. Uh yeah. Also take make it to full screen on your desktop. >> On your desktop. >> Yeah. Give me >> Chrome. Yeah. Yeah. So now take take the cursor to each cluster. Let's see. So here is Q name. Uh okay. Let's just take slow slow slow. Let's read one one at a time. >> Okay. Yeah. So it says that the makeup revolution right David Beckham Beyond man. All right. Take to another one. So these products should be identical nearby. This is vermicelli rice jelly party mix. So it's kind of has grouped these things together based on some criteria. Okay. And so on. Okay. All right. Is it based on price or what is the basis for this presentation? >> Yeah, I think the hair mask and other things the variables are above, edibles are below. So, it has basically grouped the things based on that. Yeah. Yeah. You can see that those are hair related product, >> right? >> Beauty products are clustered together. >> Oh, beauty products are clustered together. Okay. The edible products are clustered together. Similarly, it made the different clusters. >> I see. Wonderful. Wonderful. This is very useful. Okay. This is something that most of the existing tools cannot cannot achieve. A question from Anum is how could we know that the information we are getting are 100% accurate because this information is coming after running the code on your data. So if the let's say when the LLM hallucinates LLM will hallucinate in writing the code and the code will be wrong and if the code is wrong you won't get results you'll get the the the mistake in terms of syntax error or library missing or other things but there's a very low chance that it'll hallucinate and say that these are something okay and we also make sure that part yeah can Can you show the code please? >> Yeah, here this is this is a Python code and this also answers another question that was there that can it be used for uh learning data science? Yes. The reason we are keeping the code transparent to everyone is to make sure that that they know what's going on, they can learn what's going on and so on. All right. Very good questions. Thank you, Anu. Let's go ahead Glenn. >> Yeah, that's all from me. Uh, so this was the basic product recommendation. >> A question. Yeah, a question from Anum is that can we modify the code? Answer is no. You can't directly edit the code, but you can tell tell LLM to do it for you. Okay. And yeah another question very interesting question from Arun is that does it have guardrails? Of course we have really hard guard rails. There are two kind of sandboxes environment that we have designed. One sandbox for running the Python code and other sandbox or other firewall called SQL shield for interacting with your database. while interacting with your database you we can define our own rules uh on top of your existing database. Okay. So that it protects the data from malicious queries, wrong queries or mistaken queries. Okay. All right. Go ahead. >> Yeah. Uh that's all from me. So I was able to build a product recommendation system group into clusters with a simple command without any uh tech giving any technical aspects to the uh in the prompts. So it's very powerful that you can build recommendation systems without having to code and uh having to having the knowledge to do it uh but uh by just giving some prompts. Yeah. So we can continue with the next uh case study. >> All right, go ahead. Who is next? Share your screen. >> Uh hi everyone. Uh I'm Kajena uh part of the team building tunnel and uh I'll just uh walk you through the case study that I have prepared. Second. >> Yeah. Are you able to see my screen? >> A question from Arun. Anum is that can it also write the recommendation or report for us after analyzing? Yeah. Yeah. Of course, you can ask it to prepare a presentation, prepare a report of all the work or you convert the whole thing into an app. It can do all of that. All right. Go ahead. Yeah. Uh can you see my screen everyone? >> Yeah. >> Yeah. >> Okay. Great. So uh for this uh use case basically I wanted to understand I wanted to segment the customers uh that are arriving on uh my website and basically this is based on an FMCG data set right. So I asked her no to group uh the customers uh based on various uh uh reasons right like uh whether they are high spenders or uh frequently looking for their loyal customers and how much are they spending on average and few other insights. So uh so basically uh the process is quite similar. it will start uh uploading the CSV file right start understanding the data and then uh it will start processing on your query right so over here uh as you see uh turno starts to inspect the discount applied and promo code use fields to correctly flag discount usage per purchase right and uh it creates takes a code uh runs a code sees an observation and then you know comes out with a thought whether it's able to meet the uh like if it's able to solve the query right and uh it continues from there on. So over here in the next process we see that it has uh you know taken some customer metrics and it has started to define uh segment this customers on the basis of their spending and the frequency of their website visit. Right? So it has segmented the customers into four groups. The new customers, high spenders, loyal uh customers and bargain hunters. So uh and it has given us this concise uh summary about it right how many high spenders uh the count of the high spenders uh the average annual spend of uh this high uh customers right how many times are they visiting per year and what categories are they searching for generally right electronics accessories and it has also uh given us the marketing message right so when we want to target this uh groups separately, it has uh also told us how we can do that. So for this high spenders, there is exclusive VIP deals, right? And for the loyal customers, there are um yeah, the message is thank you for your loyalty, enjoy early access to new arrivals, right? And for bargain hunters again hot deals you don't want to miss. So it has uh given us this information which was very uh useful. So I ask it to dig deeper into this segment and help me understand seasonal patterns whether these customers shop differently throughout the year right uh how how do they behave during the holidays or during the sale seasons right so again uh turno analyzes the data that we have given it runs the code observes it and then uh gives us this information. Yes. as uh so this is an interactive dashboard of uh you know how uh the customers are spending uh uh during fall during each season right spring summer winters and uh yeah so it uh okay so I ask it to give me this into another visual right so in a proper detailed visual with the messaging, the top categories, the average visit and uh then a pie chart of the same. So basically you can ask turno to uh give you visuals in any format that you want right and then you can download it easily and uh this was what we wanted to show through this use case right so basically any FMC if you're into uh it it's easy to do any analysis you can do prediction you can do segmentation you can uh you know uh do forecasting Everything uh that a normal data scientist can do you can do it with uh tenoi. >> Wonderful. Wonderful. >> Yeah. So that was from my end and uh uh Tij you can go ahead. >> Uh yeah thank you Vinda. Okay. So I'll share my screen and all right. Okay. Uh so my screen is visible right? So uh the use case I worked on is basically uh customer retention analytics. So basically what actually I wanted to know when the person will buy again that's the problem a simple problem statement. Okay. So we have a data uh of uh of customers and that data uh have multiple columns knowing their age uh geographical uh reason uh purchase type uh what item they have purchased uh and uh and many more things. So this is the kind of data sheet uh we uh I took into consideration. So we have different different columns. So uh a lot of uh insight uh about the uh customer we have and on the same we want to know when this person will purchase again. So we have it review when they are purchasing what color they are choosing what sizes they are purchasing and all. So we gave this data to Turno and simply uh I wanted to know I wanted to figure out what would be the perfect timing to you know reach out to these customer when are they when they are ready to purchase. Okay. And uh I also wanted to know when uh uh how some people uh shop um every week or alternate week uh every month after few months. So all these things I wanted to know and I wrote it in a prompt and and I asked uh To to do the analysis. So uh after that uh it basically uh inspected uh district values and uh uh it mapped frequency purchase uh to each category and uh we got uh uh this detailed uh structure of uh these uh the data we have provided. So I wanted to know uh again uh so I wanted to check if turno uh to uh could give me based on the past uh you know uh shopping frequency when will the each customer probably shop again and are there any pattern do people like to uh buy again who have bought in winter uh will buy again in uh spring and uh I wanted shopping readiness score uh of each customer and which customers are probably uh ready to buy right now. So we can you know reach out to them uh at this point of time and who are those customers who will buy after a week. So or or uh or or after after a month. So accordingly we can plan you know uh campaign email campaign or uh accordingly we can also plan our inventory uh and all. So uh this is what I wanted to know. I gave it uh gave the prompt and uh it has given uh the next purchasing timings and uh uh so the people who uh so based on the customers it has given me like these these customers will buy like in these this these days less than 10 uh 7 days we can contact those person now and uh after a week we can contact those people who who uh uh uh purchase like in between uh 8 to 14 days and uh so these uh details it has given me and uh uh now so this is the data is difficult to you know understand for me so I just want to visualize it like how many candidates are there whom I should be reaching out now or how many customers I'll reach out after a week so this is basically I wanted to know so I just asked to you know uh create a visual representation of the data and test given uh maybe the data in in a visual form. So around 539 customers we can contact immediately or and uh around 1,000 candidates we can contact next week and then uh 5353 we can contact next month and after a month uh we have called it later there are 1,700 candidates. Okay. So uh an average days and season uh we have another visual representation of the same and uh this is the readiness score distribution of red uh readiness score. Okay. So uh now uh I again wanted to know uh if can turn to tell me what factors uh makes you know customers to buy again. Is it is it their age? Is it how much money they have spent last time? How was their their how are their ratings? And uh so these are the few prompts I asked to know more uh about uh you know uh so uh I asked those uh and uh it has given me uh uh top predictor uh of uh immediate readiness. So people who want to uh purchase now or within 7 days what what are the uh uh you know predictor top predict predictors are there and uh in the end uh feature importance uh expected is until uh the next purchase they spent is uh prior activities and satisfaction also play a major role. in the feature importance analysis. Uh so this all provided in detail and also uh uh in the visual representation form and at the end I asked uh to turn to you know create a report and it has given me uh detailed report that you can see here. So uh that is all uh uh from my side. Yeah. Thank you. Wonderful. Wonderful. Yeah. A question from Sur is that in the current implementation, do you have a method to uh reuse a query previously? Yes, that's exactly what app service is, right? Uh Suresh that let's say you are finished, you got the result. Now what you can do is you can create an app with parameters start date, end date or something like that and it'll be there. So you can select the the the the parameters. All right. Yeah. I I wish I would love to show you that part very soon. Probably uh in in um next couple of weeks. So it's under under uh building right now. We are working on that. Wonderful. Wonderful. This was very helpful. Uh thank you for wonderful set of questions. Everyone stay tuned with uh turno and you will be notified as soon as we are launching new uh new feature. Okay. In the products. Yeah. Go ahead. Somebody has raised their hand. Yeah. Somebody will raised their hand. Just Yeah. Uh question is who should you talk to? Just just uh drop an email to probably uh reach us at at turno.ai. Okay. Or right um one of my teammates can uh re reach out to you. Gajender can uh you know Gajinder will be available. All right. Gajender you can reach out to Sesh. Okay. All right. All right. And wonderful, wonderful, wonderful set of questions, wonderful um uh demonstrations. And we're looking forward to hear from you. Please go ahead everyone uh on the on cloud on turno website. Go to turno.ai and sign up. You'll get the free Yeah. Okay. Kajendra, you can share your number as well with Sur and yeah, there is one more demo pending uh from Rohan. Rohan, please go ahead. Sorry about I thought that all are over. I'm sorry. >> Oh yeah, it's completely okay. Okay. Hi everyone. I'm Rohandas and uh first of all, thank you everyone. you are being with us till now. I think this is too late and I hope that this is expire uh you know inspiring and more let me share my screen. Okay. Okay. Yeah. So my use case is the forecasting sales right? So the problem statement uh I was looking into is that FMCG companies needs to understand their patterns of sales right because uh they have to uh get their productions and inventory all all of those plannings right so we are having the historical sales okay from 202024 and we are looking into the key uh you know sales drivers which is channels regions categories promotions and all of these things and we are forecasting the future demand of that uh productions you know uh we are doing it for 6 months of the future and uh aligning also the we are also aligning the marketing uh strategies with this okay so we can target all of those uh you know uh required things in FMCG so for the approach I am using the data uh standard you know data approach that we are first of all preparing the data we are removing the negatives and all handling the missing datas and do a explorative data analysis looking into the trends okay and then we are uh doing this forecasting for weekly sales and we are doing it for 6 months ahead also we are getting the uh recommendation insights uh from turn so let's see it in action Okay. So, we have already done it and I'm just letting you go through it. Okay. So, my prompt was this and I have also uploaded uh data data set. Okay. For attached to this prompt. So, the prompt was I need to uh you know uh sales do the sales forecasting for FMCG company and our goal is to getting the better uh understanding of the sales pattern. Okay. So I have told him that I have attached this uh data set okay which is in CSV format and I need uh you know we have this all the columns and I need you to think of a uh data scientist and do all of the things to get a forecasting. Okay. So it came up with the thought of uh u getting the uh CSV loaded and looking into it. Okay. So it wrote this code to get the shape of the data, get the information of the data and first five rows of the data. And we can see here the observation. Okay. So the shape of the data is uh something one like 90,757 of the instances and 14 columns. And these are all the information of the data that these are the columns and these are the uh you know u instances of each column. Okay. And there is no not null in everything. So there is no missing data. Then we are having the five first five rows of the data. Okay. Then it came up with the whole approach. So whole planning of this how we will do that. Okay. You can see in the prompt I have never told him to do that and it still came up with this. So that's why we are calling it AI data scientist because it's it uh things of AI data scientist like okay. So the uh approach it came up with is uh getting the data clean okay remove the duplicates and basic outliers and all of the things then doing the explorative data analysis okay it will do the summary of statistics which is mean median and quantiles also for the different you know columns it thinking of and then plot the time series data okay according to the daily and weekly breakdown down into sale channels uh regions and categories and promotional flags and look for a uh correlation with between them also it will do the feature engineering for this and then it came up with the model for the forecasting. So it is thinking of surma and profit for the machine learning regressor and will train and evaluate it and then it will give the forecasting of the data. So it wrote the code for the first step. It begins with the uh sum of the data okay and some of the missing values all of those things we can see it in the observation as we know already we have not any missing data okay and no number of duplicate rows also okay and but it it can analyze that here are some negative okay uh data points so it came up with this that it will remove the negative data points likely be uh data errors Okay. So it write the code to remove the data uh negative data points from the data set. So after doing the processing it came up with this that rows removed due to negative uh three. Okay. And then the clean data set having this shape and we can see here the clean data set. Okay. Okay. So it came up with all of these things and then I have asked him to see into the second step which is uh showing the graphs and all to give the EDM. Okay. So I asked him to do this on the clean data set and I wanted to do the aggregate unit sold of the weekly basis and get a clear view of the sales trend. Also uh I asked them to clearly uh get a you know a visualization of a line chart to see the trajectory of the sale. It came up with the code and we can see here the graph. Okay, this is the graph. It came up with this and which is total weekly sale units. Okay. And okay, this is the line chart for this. Then I uh asked them to dig deeper into the sales drives. Okay. So we need to analyze the channels, regions and categories. So I asked him to create a bar chart of each of them and show it to me. Also look into the promotionals. Okay. So if the promotional flag is zero or one, we can look into it that is it affecting today's sales or not. Okay. So it came up with all of these things all of these uh beautiful charts that uh by the region we are getting it this much of the sales. Okay. By the category we are getting this much of the sales and uh by the channels we are getting this also from the promotional side we are having this much of sales when there is no promotions and this much of sales then there is promotion and we can see clearly that promotions helps in the sales which is normally uh sign right. Okay. So it clearly says that I have added this all the things and you can show it here. After that I asked him to uh look into this and similar heights. Okay. So because of the data all of those are similar height. I wanted to have the written over the chart that how much it is. Okay. So we can look into it and okay I have zoomed it. Let me do it rescale. Okay. So it just labeled all of these bar charts for me. Okay. Yeah. And then I came up with the next uh step which is forecasting. Okay. So I want to have a uh you know time series forecasting for this. So I asked him to build a uh model over it on the clean data set and uh basically uh prepare the data set and do all of those things to get the forecasting and please uh I have asked him to please choose a robust model like profit or cerema. It will only use according to its uh understanding of the data. Okay. And I asked him to get this for the next 6 months and also forecast visualize the forecasting of the each weekly sales and into one chart. Okay. The historical sale also the uh forecasted sales and also asked him to get the uh confidence interval. Okay. So it it came up with all of these codes use the profit model. Okay. It installs it because we do not have it pre-installed. So you install it and then do all sort of things and it takes some times but it came with came up with this and uh here you can see that this is the forecasted sale. Okay. Uh if I show you this is the forecasted sale. This is the blue line. Okay. And this is the historical sale. So this much was the historical sale but we came up with this six month of the historical sale. Also we have the uh you know uh what we say the the upper and lower boundary of it. How much it can go upper of it and how much it can go lower of it. It is forecasted all of these things with for us. Okay. And the forecasted uh you know uh the confidence is 95% which is really good uh for uh you know uh something like happening autonomously autonomously. Okay. Then uh I started asking him for the categories and all of the reasons so that it can give the more uh forecasting over the categories and reasons on all of this. So it came up with all of these things. Okay, this is weekly sales for yogurt. Okay, this is weekly sales for the region PL South. This is the weekly sales of milk. Weekly sales of ready meal. Weekly sales of the region PL North. And then I asked him to you know give this explain me the pl north and it came up with all of these explanations that what is shown in this graph and uh why the graph is you know uh having all of these curves and all of the things. So it explained me beautifully all of the things which is shown in this graph. Then I finally came up with the summarization. Okay. So I asked him to uh give me key takeaways, seasonal uh you know patterns and actionable uh recommendations inventory and production planning all of those things. Okay. And it came up with all of the things that key takeaways are strong overall growth. Okay. And top categories which are uh we which we are having okay top regions and all of these things. Then seasonal pattern uh that it uh you know um just notice that every spring and the holiday season which is late November and December it shows the peak right and also give us the actionable recommendation. So it it came up with the recommendation itself that inventory and production planning you can do the yogurt and milk for two months of April and May because it's the peak right in the quarter two. Okay. and then resource allocation and promotional strategies all of those things it came up with itself. Okay. So this was uh it and we can definitely have a better uh model but this is the uh you know small experimentation and to show you that what is the power of turnoai and what it can do for you. This is uh not only for the analytics we you but you can also do the predictive analytics with this. Okay. >> Wonderful, wonderful, wonderful. A question from Arun is one bottleneck with LM is its cost per token consumption. Does turno uh right and does turno already optimize accordingly or user have to take care of? Yes. So basically um yeah so the way we work upon is that when you are when you are using chat GPT for example and you are uploading a file that file is counted in tokens that is important part to know in our case that file is not counted in tokens so that's a big difference second when it comes to the the prompt I think it is important to give detail prompt and proper prompts. Okay. Uh instead of thinking about money and the way we uh proposed is that it's um like the whole report that Rohan has prepared must have consumed hardly $1 or $2 maximum. Okay. So it's not that costly either. If you think about the same work if you give it to a data scientist, right? It would take probably uh maybe say uh one month to get that work done and one month of cost of data scientist would be how much? Probably um minimum of uh $30,000. Okay. And and therefore therefore you save uh like $2 versus $30,000. Okay. And uh yeah so I think I think u the cost is not that high. Okay. A question from Anum is will it recommend whether we can use other graph that would be attractive or interactive rather than one we have given in the our prompt. Yeah you can tell it that uh prepare the graph according to the data and it'll figure out the best one that suits yours. Okay. Um yeah one analogy is that the smartphone is extremely smart but it consumes very high battery then it is it so yeah so the that's what I'm saying so the with respect to the token consumption the turno is pretty cheap as compared to uh a team of uh right data scientists. Okay. And here the focus always is on accuracy, right? The the cost is any day any day cheaper than all the softwares combined. If you take a look at how much people pay for very the the the BI tools, you will know that Turno will always be cheaper than all of those. Okay. So, very good point. Arun, thank you and very good point Anum. All right, wonderful, wonderful set of questions. Any other questions? We would love to answer. Thank you Gajender, Rohan, Desir and Glenn for uh good demonstrations and all right I'm looking forward to see you all u uh using turno and feel free to share your case studies and whatever you are building and yeah a question from Arun is uh you are you are your view on aentic rag or keg yeah I I think uh those are just the tools to solve problems and whatever works what you should use that so whether aentic rag or I don't know what is c a what's a full form of c context aare generation is it context yeah cuz yeah yeah so I think I think yeah Those are basically uh I would say this that use uh like test out all the approaches to solve your problem. Whatever works, whatever works. As simple as that. Okay. All right. That will be my my take on most of these things. All right. Wonderful set of questions. Everyone I'm looking forward to see you guys and have a a question from Anom is won't this be a risk for a data scientist job? No, not really. You will see that the data scientists who understand the the analytics who understand uh understand uh models they would be they would become 10 times more more performant in in today's world there's a lack of data scientist right and essentially if you go to say a company FMCG company most of our customers Right? None of them have been able to like these are rich companies. They have billions of dollars of revenue but they have not been able to retain or hire data scientists in general. And and the reason is that that the the doing the data science job is a is a costly affair and it requires a lot of knowledge and therefore therefore we are essentially uh making the data scientist much more powerful. Like you can take the analogy, you take the analogy like when earlier when there was no motor vehicle, people were pulling the cart manually, right? And now those people who were pulling the cart manually, they when they were given the the cars to drive, they started becoming very good to drivers, right? Similarly, the data scientists those who were writing code by themselves now they would be using uh tool like turno to become 10x in their job or maybe 100x in their job. Okay. So essentially it's about empowering the industry. It's about empowering the data scientist. It's about empowering the uh the the companies to do operation excellence to do build better products to build products faster and and reduce the cost of the products. So that's the whole idea and yes so thank you everyone have a good day. Bye-bye. Have a good day. >> Thank you everyone. >> Yeah. Thank you. Thank you.