Introduction: Exploring AI in Action with WSBI

My name is Matteo Marchione, I'm the IT lead at P &O Innovation, I'm here with my colleague

Silvia and we have prepared for you this presentation, Exploring AI in Action, a real -world R &D user

journey using WSBI, which basically means we are going to see how we did embed AI into

our platform from an end -user perspective.

WSBI Platform Overview

And a few words about the platform itself, WSBI again.

again.

We started developing WSBI in 2015 by collecting together several sources and

providing a unique point of access for information which are relevant for open innovation.

So

Unifying open-innovation data sources

we have funding opportunities, funded projects, academic papers, patents, collaborations,

and the idea was again to provide a single point of access for things like think about

a competitor analysis maybe you want to run a competitor analysis and you are

interested in a specific field so you have a searching engine and you can use

keywords to create the database then you get all the academic papers patents

which relate to a topic and so you can identify which are your competitors and

benchmark and see how do you position with respect to them so it's not just

Connecting data and adding an analytics layer

about ingesting the data it's also about merging the data and creating connection

and on top of that we have then built an analytics platform which basically can

be seen as a data visualization layer so the idea is when you query the database

that you not only get the items but also aggregated information so you can have

like a map that shows you where the projects are mainly worked and topics

classification, trends, basically ready to be exploited information if you want to

run tasks like a state -of -the -art report.

Let's say you want to build a

state -of -the -art report, you know that there are projects but you maybe need a

chart which is the evolution over time and you do already have these

features out of the box.

From Advanced Analytics to an AI-Powered Platform

By following the pathway of the advanced analytics now

WISB is turned into an AI powered platform which basically means we are

and with lots of AI to his research understanding the context the contents

and also building end -to -end flow which speed up a bit the work of the

Case study: PET glycolysis for chemical recycling

researchers but let's see what we have today we prepared a case study which is

about advanced techniques in pet glycolysis I'm not an engineer and not a

chemical engineer and a mathematician so please be patient if you have that kind

of background but in short it is about breaking down a plastic product like a

bottle of water into its raw components in order to use these raw components to

build a new plastic product so this is about plastic recyclment something we

are probably and hopefully or all confident about it's about having a

better world and well probably something which is easy to understand for all of

us and it's a real use case so what I mean by real use case we do usually

interact with end users of the platform which are to understand which are the

challenges mainly and this was a trigger for a lot of these features that we

we develop, that's also why I choose this.

End-to-End AI Workflow in WSBI

What the demo will cover

We will see an assistant or an agent as

you want to call it to make technological discovery within the

platform so we will have an LLM embedded into the application and then based on

the outcome provided by the LLM we will choose how to create the database, we

will do it in natural language and we will see that in the back -end this is

translated into a complex query.

And then we will use AI again to understand the

specific contents.

It will be a very complex patent and we will have this

smart summary features to understand which is its contents.

And then finally

what it is most interesting for me, the innovation evaluation tool, which is a

tool which leverages the data of the platform to assess the innovativity

potential of a project idea and it leverages AI as a natural interface with

the human and several data -driven tools like machine learning tools and we will

we will see that.

Product Walkthrough

Landing page and datasets

This is the landing page of WSBI.

In here we can see all the

the datasets that I have mentioned, the patents, the papers, projects, and so on.

WSBI Agent: conversational technology discovery

Let's start from the WSBI agent.

So here we are conversating with an LLM and we can use it for making technology discovery.

So let's say I'm going to ask something, which, okay, pretty easy question.

Why it is important to have an AI embedded into an application just, you know, for not

doing like back and forth with the application in Google, the application in ChargeGPT, maybe you

don't have a ChargeGPT license and you want a reliable LLM you want to speak with.

Again,

this was one of the challenges that were posed by our end users and this is simply integrated here.

Of course, this step usually requires several iterations, so you want to ask something,

then you want to deep dive.

This is something we are not going to do today.

Let's say we find an argument which is interesting for us, which is microwave

assisted glycolysis in this case.

Natural-language search over patents

We move to the specific data sets in which we

are interested in, which in this case is the patents as we have said, and we want

to query the database.

We are probably not database experts and we can use

this natural language search capability.

The reason why I did add pet there is

because pet is an ambiguous word, so usually this, in the past at least, this

was giving results about like the animal world, like research

about veterinary thing.

And we also have microwave which is an

ambiguous word which could lead to things which relate, you know, house

devices and this kind of thing.

So what we want to see here, then if we run this

query, the system in the backend is translating it into something which is a

bit more complex.

How the backend validates and enriches queries

What we have in the backend here is a genetic pipeline which

basically understands the concept and first of all it validates that it is not

like a malicious or injection attempt.

Then it is understanding the context of

the research and it is extracting the main concept, then enriching it with

synonyms but why this is interesting because as we can see by analyzing

quickly the result everything is now related to chemical chemical recycling

chemical recycling chemical recycling so basically why we are embedding AI in it

because usually I mean basically to democratize the access to the data by

this platform so not everyone is very confident with database

query and in this way we can simply use natural language to ask things to the

database and at this stage once we have identified the context of a research and

Understanding complex documents with smart summaries

we target the right contents maybe we want to open a content again we are

checking patents and yeah probably we have to go through the complex text of a

patent which is usually long and complex again.

So what we added here is to have

again an AI to generate a smart summary which is I mean summary of the content

itself providing a structured information which gives yeah a

description which are the main claims which is the problem that the patents

target to address and well at this stage usually again this is an iterative

process so you want to understand which is the context check check patents check

papers check projects at some point you come back to your team and you maybe

have an idea about how to move forward with respect to this patent or in

Innovation Evaluation Workflow

in general to the state of the art and you're not really sure that your idea is a good one

in terms of eligibility and what we are working now is a workflow which I have run previously

since it takes like two or three minutes and I was sure we didn't have the time for doing

Scoring innovativeness with auditable platform data

it but basically we have this section workflow in which you can access this innovation evaluation

tool and then you insert the project idea you press run and after three

minutes you get an assessment the assessment is based on the data of the

platform and it basically compare your idea with benchmark of projects papers

and it provides an overall score of

innovativity and a radar chart which

displays the different drivers.

Basically for us something it is

innovative if it is original.

So this

novelty drivers basically measures how

much this idea is different from what we

have in the database and if it is

aligned with or better if the topics of this idea describe increasing or

decreasing trends and if the idea is useful what does it mean useful for us

is if there are properly budgeted funded opportunities for which suits for

this idea and which is the impacts.

What the LLM is doing here is basically

providing just the natural language interface, so the analysis itself is not

performed by the LLM.

We have other analytics tools which leverage machine

learning or statistics to get the numbers and then the LLM is simply

describing which is the outcome of the analysis to the user.

And let's see how a

Driver example: originality via similarity search

couple of drivers works and then we move ahead so how to assess if something it

is original we perform a similarity search against the database and we

retrieve which are the most similar projects papers and patents we have the

similarity is computed by using vectorizers models and we are basically

mixing the information about how much something is similar or dissimilar to

idea with the year in which the project so basically how much time it passed

since this idea was released and this benchmark set basically defines all

these scores for the projects and the average score of the benchmark set is

the score that our idea is getting so in this example we have 40 so it is not

that innovative because it seems to be innovative when comparing it to the

patents but definitely we have a lot of projects in the same area so this is

probably something that we want to to improve and iterate again the process

study I've described before.

Driver example: trend alignment

Trend alignment, again the question here is if

this idea is aligned with trends which are increasing and again this is

computed by benchmarking data in the database.

So in here we can see which is

the distinguish between the data itself which comes from the database, those are

reliable, this is auditable and I mean basically trusted data and what the LLM

is doing is just explaining the content of the data so it's natural language

interface with the data itself and let's switch to the maybe I mean just to the

last part so once we have understood that our idea is maybe original it is

Driver example: funding-fit suggestions

is probably aligned with increasing trends, we are searching for funding opportunities.

Then we switch to this third tab in which the system is suggesting codes which could

fund this kind of activity, and of course these are also part of the database, so the

user can select it and then check if this is aligned with the scope of the projects.

And well, this is all that I wanted to share, trying to get a very quick summary.

Conclusion: Where AI Adds Value in WSBI

We used AI in several ways, starting from an assistant, so again, having a sparring

partner to conversate about technologies then identify a topic of research and

using AI to properly query the database then when we analyze a specific

content we are interested in understanding it better and using AI in

it and then finally we have these big end -to -end workflows which leverage

machine learning, AI and I mean analytics techniques to provide

explainable information.

I know it was a lot in a very short time, I hope it

was clear and yeah please.