Introduction

So, yeah, I founded Eli, sorry, okay.

Founding and early traction

I founded Eli out of my PhD at the University of Roma 3 nearby.

Eli raised around 10 million euros in venture capital so far. And we brought our product to more than 100 companies in 15 countries in the world. 80 % of our customers are from abroad. We have right now more than 10 ,000 AI predictive model up and running to our platform.

Enterprise customers and market presence

And we have some logos here of our customers. We're talking about big companies. So I'm going to talk about enterprise AI companies like, you know, Intesa San Paolo in Italy, but also MetLife.

This is a large insurance company based in the US. US, you know, UNICEF, Orange Telecom is the largest telco in France.

Leadership team additions

Last year, some executive joined us like Antonio Sciuto, he's the CEO of our company now. Since you know, last year, he was the global growth officer at Salesforce in the US, working with Mark Benioff.

And the CEO of our North American companies, Kirsten Moody, that was chief data officer for Prudential Finance and State Farm which is the largest insurance company in the US.

The enterprise AI gap: adoption vs. measurable impact

So let's talk a little bit about enterprise AI right now.

If you look at BCG and McKinsey state -of -the -art on AI you can see that the global AI spending for corporations and enterprises is a lot of money and we We have an 88 % of AI adoption, so they are adopting it everywhere.

But if you look at impact on the P &L of companies, only 6 % of the companies were able to measure an impact in terms of increasing of revenues or reducing costs to more than 55 % of EBIT. And 95 % of Gen -AI projects didn't impact the P &L. Wow, why is this?

What AI winners do differently

Focus on predictive AI and core operations

because you know if you look at how winners in this space are tackling this topic they focus on three things the first one is they focus not only on Gen AI but even on predictive AI because you know predictive models has always been

there as always delivered P &L impact they focus on core use cases so not just you know optimizing something in the back office or stuff like that but they really focus on the core operation of the company and they rewire the whole

process around the AI they don't do just you know a small optimization and our

A vision of models embedded across every business function

vision is that for each you know for any given company any given business function from HR to marketing from customer service to finance to supply

chain we envision a world in which every companies has you know hundreds of predictive models that guys decisions because companies right now have a lot of data so they could potentially build a lot of predictive models you know

Common high-value predictive use cases

machine learning deep learning models anything from you know lead scoring scoring, clustering of the customer base, churn propensity, cross -sell up -sell, next

best action, ticket prioritization or clustering, fraud detection, dynamic pricing, predictive maintenance for manufacturing companies, or demand forecasting for any given retailer or CBG company or stuff like that.

Why predictive AI doesn’t scale in most companies

But usually this is not the case. So usually companies tend to have a very limited adoption of predictive AI. We talked to a lot of big companies and they have like zero to five predictive models up and running.

Why?

The people and platform work behind a single model

Because to build just one of these models here, it's a lot of work. You need to have heavy teams of data scientists, AI engineer, data engineer, IT, DevOps

hopes that will take real dirty data scattered across, you know, companies, data sources, the CRM, the ERP, data lake, data warehouse.

And those guys needs to work on data integration, data cleaning, feature engineering, data transformation.

And after that they need to do AI modeling and backtesting and deployment and integration, and this is all just for one predictive model in production.

Long delivery cycles and continuous retraining

McKinsey, again, says that this process usually takes six to nine months to take a predictive model and to put it into production for a large company with a lot of data. And when this process is over, it's not over yet, because after six months, the model will be outdated because the data will change. You need to retrain the model over time.

So it's a lot of work. And if you want to take your company to, I don't know, 10 ,000 predictive models or 1 ,000 predictive models or 100 predictive models, it's not scalable. You can't do it.

Eli: an agentic platform for the full predictive AI lifecycle

So what Eli is, is an agentic platform that delivers predictive models into production and will take care of all the lifecycle of predictive AI, AI, starting from real dirty, messy data scattered across different sources and the EDA phase and understanding of the data and the decision on how this data needs to be transformed and cleaned and turn into a clean data model that will be able to be used to perform AI modeling

and backtesting and then to judge the model is this model good there is overfitting or data drift or whatever do I need to retrain it once a year once a month once a week so all these decisions we are talking about hundreds of

decisions talking about dividing the work in micro tasks so you need something that will take this whole process divided into hundreds of micro microtasks and actually perform it.

User workflow: choose a model, provide data, define the target

So from a user point of view, actually, it's very simple because you need to decide which kind of predictive model do you want. You can decide it through the UI or you can type a prompt, obviously.

You could choose, you know, a probability scoring model or a forecast time series model or a clustering model or a classifier or a commander and then provide the data that you want to use as a training data set and define the target variable either

labeling it or defining it with a prompt and then Eli will start the whole work of understanding the data and how each column needs to be cleaned and transformed in a way that will be able to train a predictive model and then take it into production and maintaining over time

Example implementation: an S&OP decision engine for CPG

Okay, so to show you an example of implementation of this kind of things, there is this SNOP decision engine that we have deployed to several customers around the world. And you can see here the complexity of this kind of work.

So let's say that you are a CPG company like, I don't know, Procter & Gamble or, you know, McCormick, Unilever, Mondelez, whatever.

You produce a lot of products and you buy a lot of raw materials in order to produce these products.

And you sell like 10 ,000 different products in 80 countries or you sell 100 ,000 products in 80 countries or whatever.

So we start from the

Step 1: forecasting raw material prices

the raw material price forecasting.

That means that if the company buys 200 raw materials, let's say coffee, black pepper, I don't know, whatever, oil, okay?

You will input into Eli commodity prices and macroeconomic information from the OECD and weather data and supply and demand data.

And you will ask Eli to build a predictive model model for the raw material price for the next six months, like how the black pepper price will be in six months of time.

And this for 200 different raw materials, that means 200 different predictive models.

Step 2: demand forecasting by SKU and scenario planning

Then you have from there the demand forecasting by SKU.

Maybe you have a thousand products, so you need to have a thousand predictive models that will take sales data and promo calendar and pricing structure and raw material price and you will use it to forecast sales of each one of your products each one of your SKUs for the

next 12 months but these sales forecasts will depend on external factor like seasonality and raw material price and whatever but also an internal factor like the promo that I'm going to apply to my products or the pricing that I'm

going to apply that means that from there i can do scenario about pricing and promo you know pricing and promo scenario simulation so i can use these models to start providing different scenarios in terms of promos or pricing and select the best ones so the ones that will maximize my my sales or maximize my revenues and from there you can take decisions on

From forecasts to S&OP decisions

your S &OP process your sales and operation process so we have a lot of different customer interviews that you can see on our website now I'm gonna

Product demo: building a demand forecasting model

show you a demo so let's make a very very quick example just to show you guys

let's say that I want to build I have to type demand forecasting by SKU okay so I could type a prompt to Eli or I can simply say that I want to build demand forecasting by SKU and choose the forecast model type.

Obviously forecast models are models that will forecast a time series over time.

So you can create the model, afterwards it will ask to train the model by providing data, you know, past data about the phenomenon.

Connecting data sources and assembling the training dataset

I can use several connectors to my company's data sources, you know, again we are into some enterprise environment so maybe the company has systems like Salesforce or Oracle or snowflake data breaks or whatever and they can connect their system to Eli and use it to import data I will do it

through my through a simple CSV file just to show you but it's the same through connectors and for example I will forecast sales of my products by using a forecast model so I'm gonna provide some sales data about my SKUs my products for the last I don't know three years three years worth of data

and here is an example of a file I can see here a product ID maybe I have a 100 products and basically I will ask Eli to build one model for each one of my products by labeling this column as an item ID and it will create one predictive model for each one of those

then labeling this column as a timestamp for my model and this one as a target variable then I have a lot of variables obviously in my data set that I want to use in correlation with the target function and then I can start adding new data to this data set

like for example I can add the data set contained in another data source that contains for example OECD information so macroeconomic information about economies like this was about a Brazilian economy so the GDP the the interest rates, the inflation, you know, indicators about the economy that I want to use in correlation with this table here.

And I can go, you know, on and on adding new data sources. And at the end, I will have here a data model made up, you know, in real world, you will have here like, I don't know, 40 different data sets connected together through, for example, the date.

Configuring the forecast and training at enterprise scale

so i'm gonna connect this data and go continue and choose how i want the model for example i want the prediction to be weekly i want to i want it for the next 10 weeks ahead and you know for example the model will run weekly itself and then i will launch the model so obviously this is on

demo data but you need to think that in real world use cases you will have a lot of information like 40 data sets, thousands of columns, billions of rows and you need to ask to Eli to develop one model for each one of your products and at the end of the day

Interpreting results and running promo/pricing scenarios

for example it will deliver a model like this one as soon as the model is trained I can actually see the performances in terms of accuracy of the model for example this is for a specific SKU so a specific product of my my company and I can see the first two years has been used as a training set

the last two years has been used as a test set and I can see that starting from this week the model has been used by Eli to predict sales weekly for the next 18 months and the purple line predicted the sales to go down and up and down and up and slowly down at the end up and in the blue line I can see the real data so the real outcome of sales and with a forecast accuracy of 86 percent in this case

Eli was able to predict this information so what's important here is that for example when Eli predicts when the model that Eli built predicts a spike in the sales maybe that is because I said to the model that I want to that I want to apply a specific promo or pricing decrease for this period of time that means that if I'm going around several predictions

several scenarios the prediction will change and that means that I can run like I don't know 10 different pricing and promo scenario simulation and decide which one will deliver the most value in terms of P &L impact to my company and this will be done for each one of the products each one of the geographies and and whatever that I handle.